搜索结果：sentences

In sentences such as "John remembered the boy took some time to rest", the locally ambiguous noun phrase (NP) "the boy" is initially parsed as the direct object of the matrix verb "remembered" (the object analysis). When the embedded verb "took" is encountered, the NP is revised as the subject of the embedded clause (the subject analysis). Open questions are how the real-time resolution of this complement ambiguity is influenced by semantic/categorial constraints (i.e., whether the locally ambiguous NP is a semantically/categorially appropriate object of the matrix verb) and selectional frequency (i.e., the frequency with which the matrix verb takes a direct object NP). The present study addressed these questions and also examined whether temporal adjuncts, which can bias parsing towards the object analysis (e.g., "John remembered the boy after…"), influence real-time ambiguity resolution. The results showed that semantic/categorial constraints and selectional frequency drive the processor towards the subject analysis before embedded-verb disambiguation. However, temporal adjuncts gradually increased in influence and ultimately overrode these biases during processing. The observed parsing process was also simulated within an interactive constraint-based framework. Together, the experimental and simulation results suggest that real-time sentence processing is dynamically shaped by multiple competing biases.

Integrating auditory and visual information can improve intelligibility and neural tracking of the speech envelope. Speech-relevant tactile information also improves tracking, yet effects on intelligibility are mixed, possibly due to no regular exposure to speech-relevant tactile information. We used short-term audio-tactile training to advance understanding of audio-tactile integration during speech perception. 64 younger adult participants (aged 18-29) took part over five days, during which electroencephalography was recorded. Participants completed a speech-in-noise task, with audio-tactile and audio-only stimuli. They then received training with either tactile information that was congruent with sentences heard (trained group) or incongruent (pseudo-trained group). After three training sessions, participants completed the speech-in-noise task again. Two weeks later, participants returned for a follow-up. Effect of session (pre- or post-training) on speech intelligibility was significant, with no significant effect of group (trained, or pseudo-trained) or stimulus (audio-only, or audio-tactile). Before training, there was a significant increase in speech-envelope tracking accuracy with audio-tactile stimuli, suggesting enhanced neural representation of audio-tactile stimuli. However, there was no benefit of congruent training to audio-tactile tracking. There was an enhancement of audio-only speech-envelope tracking following incongruent training. This suggests that speech intelligibility and tracking of audio-tactile speech are not enhanced by short-term training. This work challenges prevailing assumptions by demonstrating increased speech-envelope tracking accuracy that was not linked to enhanced speech intelligibility. However, these findings are limited to short-term, bottom-up audio-tactile training. Alternate training approaches and further controls, such as tactile-only sentences or audio-only training groups, should be explored in future research.

Neural activity has been shown to track hierarchical linguistic units in connected speech, and these responses are modulated by changes in speech intelligibility resulting from spectral degradation. In this study, we manipulated prior knowledge to enhance the intelligibility of physically identical speech sentences and tested whether this improvement would strengthen neural tracking responses. Cortical magnetoencephalography responses were recorded from 23 normal-hearing participants while they listened to intelligible speech followed by either the same (matched) or different (unmatched) unintelligible speech. When prior knowledge was available, cortical coherence at higher-order linguistic rates, particularly phrase and sentence rates, was enhanced relative to the unmatched condition and was predominantly lateralized to the left hemisphere. In contrast, cortical coherence to word-level units, which aligned with acoustic onsets, was bilateral and did not show a significant modulation by contextual information. No such coherence enhancement was observed when unintelligible speech preceded intelligible speech. This dissociation suggests that cerebral tracking of linguistic information is directly influenced by intelligibility, which itself is strongly shaped by physical speech cues. These findings provide an objective and sensitive neural index of speech intelligibility and help explain why previous studies have reported no effect of prior knowledge on cortical entrainment.

This mini review surveys published work on automatic summarization of legal judgments. We focus on how natural language processing handles argument structure, discourse, and the role of citations and precedent when systems generate shorter versions of a case. The field began with extractive methods and classical machine learning, and moved through graph-based and neural models, and to recent transformer architectures and large language models (LLMs). Some of the recent papers pair LLMs with retrieval augmentation so that outputs are faithful to the source text. A scheme is repeated in the domain that whether researchers should work at sentence or clause level, how rhetorical roles are defined, and how to judge a summary when fluent wording can still misrepresent the judgment. Many limitations still exist where many studies rely on general-domain setups or small, jurisdiction-specific corpora; citations are often treated as meta-data rather than part of the argument; and evaluation still depends heavily on overlap with reference summaries, which only partly reflects legal quality. It is also non-practical to compare results when datasets and benchmarks differ from one country to another. We situate this review next to wider surveys of legal summarization and with LLMs in law, but with only the judgment documents. Finally, we guide in future directions with more languages and jurisdictions, summaries that users can trace back to the judgment, and evaluation that mixes automatic scores with expert review where it is practical.

Verifying that a proposed solution truly resolves a scientific problem is central to trustworthy reasoning and retrieval. Using SCP-116K, we build 177,836 balanced problem-solution pairs (88,918 matched, 88,918 mismatched) spanning diverse STEM disciplines, and frame verification, following TRIZ/IDM, as distinguishing matched from mismatched pairs. Comparing lexical, retrieval-style, and lightweight neural models, our best model (RoBERTa + Slim ResNet, frozen sentence embeddings scored by a residual MLP) reaches an AUC of 0.966, an F1 of 0.905, and a LogLoss of 0.238. A CPU-friendly TF-IDF + Cosine + Elastic-Net baseline trails by 1.6-1.7 AUC points yet runs roughly 250× faster in about 1.5 GB of RAM, a strong efficiency-accuracy trade-off. The probabilities act as re-ranking scores over candidate solutions; we read the high ROC-AUC as pairwise discrimination and absolute accuracy as an upper bound given the synthetic negatives.

Few large-scale, semantically informed mappings of medical education research exist: traditional methods such as keyword co-occurrence and Latent Dirichlet Allocation cannot capture the contextual meaning of scientific abstracts. This study aimed to identify major research topics in medical education, classify their temporal trends, explore temporal associations with major policy and technology developments, and characterize their structural positioning. We applied BERTopic, a transformer-based topic modeling framework, to 276,253 English-language publications from PubMed, Web of Science, Scopus, and OpenAlex (2000-2024). Documents were embedded with a sentence-transformer language model and clustered into topics; labels were independently reviewed by two authors. Trends in each topic's relative share were classified using Poisson regression with an offset for total annual output, autocorrelation-robust Newey-West standard errors, and false discovery rate correction. Topic positioning used an embedding-based centrality-density heuristic. BERTopic identified 39 topics (7.0% outliers); Topic 0 was a residual, non-substantive cluster excluded from interpretation. Among the 38 substantive topics, 10 showed significantly increasing relative shares (Hot; including AI/ChatGPT, virtual reality simulation, burnout and wellbeing, diversity and equity, empathy, and climate and health education), 12 showed decreasing shares (Cold), and 16 were Stable. Classifications agreed across alternative standard-error specifications for 35 of 38 topics; simulation-based education, clinical imaging, and interprofessional education were significant only under less conservative errors. Temporal associations with policy and technology reference periods were exploratory and not causal. The positional map distinguished established themes from emerging, terminologically unsettled niches. The relative emphasis of medical education research has gradually shifted toward technology-enhanced and socially responsive themes, while foundational topics continue to grow in absolute volume. These findings offer one data-driven, assumption-dependent lens on the field's evolution to inform curriculum planning, research prioritization, and policy.

The Indian Lok Sabha generates a continuously expanding corpus of legislative records, predominantly archived as unstructured PDF files. Effective public access remains limited due to the shortcomings of keyword-based retrieval systems and the hallucination risks of general-purpose Large Language Models (LLMs). This paper presents a domain-specific, resource-efficient Retrieval-Augmented Generation (RAG) framework employing DistilGPT-2 (82M parameters) as the generative model, grounded via FAISS-based semantic retrieval using Sentence-BERT embeddings. The pipeline integrates multi-stage PDF preprocessing, semantic indexing, and context-aware response generation. Evaluation was conducted on 450 queries spanning simple, complex, and compound categories, assessed by human annotators using factual accuracy and a five-point relevance scale. The proposed RAG + DistilGPT-2 framework achieves 94% factual accuracy and a relevance score of 4.6 out of 5, substantially outperforming zero-shot baselines (80% factual accuracy without RAG), while maintaining an average end-to-end inference latency of 1,800 milliseconds (ms) on standard CPU hardware. The results demonstrate that combining domain-specific retrieval with a lightweight generative model effectively mitigates hallucination and reduces computational overhead, offering a scalable, transparent solution for e-governance applications without reliance on GPU infrastructure.

Although Long-Term Care (LTC) benefits aging populations, evidence on government-supported LTC for socioeconomically disadvantaged older adults in low- and middle-income countries remains limited. This study examines how government-supported institutional LTC is associated with physical and mental health outcomes among welfare-dependent older adults in rural China. A cross-sectional analysis was conducted among 313 welfare recipients aged 60 and above in Shaanxi Province. Health outcomes were assessed using validated measures of physical health (self-rated health, pain, functional limitations) and mental health (depression, anxiety, loneliness). Regression models and propensity score matching were used to examine associations, and potential mechanisms were further explored. Government-supported institutional LTC was significantly associated with better physical and mental health, particularly higher self-rated health (ATT&#x2009;=&#x2009;0.177, p&#x2009;=&#x2009;0.005; &#x3b2;&#x2009;=&#x2009;0.079, p&#x2009;=&#x2009;0.006) and lower anxiety (ATT=-0.288, p&#x2009;<&#x2009;0.001; &#x3b2;=-0.146, p&#x2009;=&#x2009;0.016). Longer stays were associated with more favorable health patterns. Mechanism analyses suggested that these associations may relate to age-friendly environments, improved healthcare access, enhanced social participation, and better nutritional conditions. Heterogeneity analyses indicated the associations were stronger among older, illiterate individuals and those with family guardians, particularly male guardians. This study provides robust empirical evidence of associations between government-supported institutional LTC and better health outcomes among socioeconomically disadvantaged older adults, highlighting its potential role in enhancing well-being in rural settings. These findings emphasize the importance of expanding equitable, integrated, and high-quality LTC services to support sustainable, person-centered aging care for vulnerable populations.

This study aimed to compare the effects of six weeks of Kinesio taping with treadmill training, Mulligan taping with treadmill training, and treadmill training alone on ankle dorsiflexion passive range of motion, balance, ankle stability, and fall efficacy in individuals with chronic stroke. Thirty-three individuals with chronic stroke were randomly assigned to three groups (n&#xa0;=&#xa0;11 each). Interventions were applied for six weeks. Outcome measures included ankle dorsiflexion passive range of motion, static balance, dynamic balance, ankle stability, and fall efficacy, assessed before and after the intervention. Between-group differences were analyzed using one-way analysis of variance. Significant group differences were observed in ankle dorsiflexion passive range of motion, eyes-open static balance, dynamic balance, and ankle stability (p&#xa0;<&#xa0;.05). The Mulligan taping with treadmill training group showed greater improvement in ankle dorsiflexion passive range of motion than the treadmill-only group. The Mulligan taping with treadmill training and treadmill-only groups improved eyes-open static balance more than the Kinesio taping with treadmill training group. Both taping groups improved dynamic balance more than treadmill training alone. Ankle stability improved most in the Mulligan taping with treadmill training group. No significant differences were found in eyes-closed static balance or fall efficacy. Mulligan taping with treadmill training may improve ankle dorsiflexion range of motion and stability, while both taping interventions may enhance dynamic balance. Treadmill-based gait practice may contribute to eyes-open static balance improvement.

Endocytosis replenishes synaptic vesicle (SV) pools that are required for persistent transmission of chronic pain signals within nociceptive spinal circuits. The nociceptor-specific contribution of SV endocytosis to pain and the therapeutic potential of endocytosis inhibitors are unclear. We identified SV endocytosis in nociceptors as a critical driver of ongoing pain and developed a gene-based strategy to target this mechanism. Nociceptor-specific adeno-associated virus-mediated knockdown of adaptor-associated kinase 1 (AAK1) or dynamin 1 (Dnm1) in dorsal root ganglia Nav1.8-positive neurons inhibited postoperative and neuropathic hypersensitivity without affecting baseline mechanical or thermal sensitivity, locomotion or spontaneous behavior. Electrophysiological recordings from spinal neurons combined with optogenetic activation of nociceptor afferents showed that AAK1 or Dnm1 downregulation blocked the sustained synaptic transmission between nociceptors and dorsal horn neurons by disrupting SV recycling and reducing neurotransmitter release probability. Lipid nanoparticle (LNP)-encapsulated CRISPR/dCas9-repressor mRNA constructs (dCas9-R) were engineered to achieve sustained and reversible transcriptional and epigenetic repression of Aak1 or Dnm1 following intrathecal delivery. LNP-mediated gene modulation produced sustained downregulation of Aak1 or Dnm1 mRNA in sensory neurons and resulted in robust and long-lasting analgesia in preclinical models of postoperative, inflammatory, neuropathic and osteoarthritis pain without impairing acute nociception or locomotor activity. Mechanistically, targeting endocytic machinery disrupted SV recycling at nociceptor terminals, thereby reducing excitatory neurotransmission within spinal pain circuits. Together, these findings establish presynaptic endocytic regulation as a convergent mechanism underlying chronic pain and demonstrate the translational potential of LNP-delivered CRISPR/dCas9-R as a durable, non-opioid pain therapy that surmounts inherent redundancy of pain signaling mechanisms. Synaptic vesicle endocytosis in nociceptors is a critical mechanism driving ongoing pain and targeting this process with intrathecal LNP-delivered CRISPR/dCas9-mediated gene repression produces durable, non-opioid analgesia across multiple chronic pain models.

Neuroinflammation mediated by reactive microgliosis is a central driver of Parkinson's disease (PD) pathogenesis. This inflammatory process unfolds years before clinical symptoms, creating an opportunity for early intervention. In vivo imaging technologies that could detect and quantify microglial reactivity are therefore essential for early diagnosis, patient stratification, and evaluating emerging immunomodulatory therapies that target this fundamental driver of PD progression. Yet no standardized, sensitive, and specific technology currently achieves this goal. Molecular magnetic resonance imaging (mMRI) is uniquely suitable to address this problem because it integrates inherent high spatial resolution and soft-tissue contrast of conventional MRI with molecularly targeted contrast agents, enabling simultaneous acquisition of anatomical detail and functional/biological information at submillimeter isotropic resolution. Here we present a novel mMRI probe designed to specifically target colony stimulating factor-1 receptor, expressed primarily on microglia in the brain. In silico data show that the targeting ligand binds the extracellular Ig domain of the receptor. In vitro cell uptake studies with both murine and human microglia cell lines show that the probe binds the receptor triggering active cell uptake and in vivo MRI enabled effective separation of the A53T mouse model of Parkinson's disease from control mice using radiomics-assisted MR image analysis. Ex-vivo immunohistochemical analysis showed signal from the probe largely in the cytosolic compartment of IBA-1 reactive cells, confirming that the observed in vivo MRI signal is due primarily to retention of the agent by microglia. This novel technology has the potential to interrogate the rgional presentation of microglial activation in PD. A microglia targeted MRI probe generates disease-specific contrast after injection, clearly distinguishing A53T Parkinsonian mice from controls.

As we consider entrusting large language models (LLMs) with key societal and decision-making roles, measuring their alignment with human cognition becomes critical. This requires methods that can assess how these systems represent information and facilitate comparisons with human understanding across diverse tasks. To meet this need, we adapted representational similarity analysis (RSA), using pairwise ratings to help quantify alignment between AIs and humans. Among the models we studied, GPT-5-mini and Claude Sonnet 4.5 showed the strongest alignment with human text ratings. Llama-4 was the best aligned open-source model. However, gaps between LLM and human behavior remain. No model we studied adequately captured the inter-individual variability observed among human participants, and models only moderately aligned with individual human responses. We demonstrate the utility of this approach across multiple modalities (words, sentences, and images), helping further our understanding of how LLMs encode knowledge, and enabling an examination of alignment with human cognition.

Acinetobacter baumannii (A. baumannii) has become a serious clinical threat due to its increasing antibiotic resistance, particularly in hospital environments. As conventional treatments become less effective, there is an urgent need to explore alternative therapeutic strategies. One promising target is the histidine biosynthesis pathway, which is absent in humans but essential for bacterial survival. This study focuses on ATP phosphoribosyltransferase (HisG), the enzyme that catalyzes the first committed step in histidine biosynthesis. We investigated whether Bleomycin, a known anticancer drug, could interact with and potentially inhibit this enzyme in A. baumannii. Initial virtual screening identified Bleomycin as a top candidate based on its favorable docking score. Follow-up molecular interaction analysis revealed strong binding within the enzyme's active site, involving key residues critical for catalysis. To validate this computational prediction, surface plasmon resonance (SPR) was used to assess the binding kinetics of Bleomycin to recombinant HisG. The observed dissociation constant (KD&#x2009;=&#x2009;270&#xa0;nM) indicated moderate to strong affinity. Further, in vitro antibacterial assays demonstrated that Bleomycin significantly inhibited A. baumannii growth, with a minimum inhibitory concentration (MIC) of 7.8125&#xa0;&#xb5;g/mL. Time-dependent growth studies revealed strong bacteriostatic activity of Bleomycin at and above MIC concentrations. Overall, the results suggest that Bleomycin strongly interacts with the HisG active site and suppresses the growth of A. baumannii, highlighting its potential as a promising HisG-associated antibacterial agent. Further in vivo work is needed to assess safety and efficacy; this study opens the door to new therapeutic options using existing drugs to tackle multidrug-resistant pathogens.

Regulatory CD8 + T-cells (CD8 + Treg) are a distinct yet understudied T-cell subset capable of simultaneous immunosuppression and cytolysis. Here, we characterized induced human CD8 + Treg (CD8-iTreg) generated from peripheral blood CD8 + CD25&#x207b; T-cells using anti-CD3e mAb-loaded artificial antigen presenting cells, IL-2, TGF&#x3b2;, and Rapamycin. These CD8-iTreg differentiated into a stable, highly proliferative bifunctional population with suppressive activity comparable to CD4-iTreg while retaining cytolytic capacity similar to conventional CD8&#x207a; cytotoxic T lymphocytes (CTL). Multi-parameter spectral flow cytometry and single-cell RNA-seq revealed a distinct immunoregulatory signature: a predominantly Treg-like profile marked by tissue-residency marker CD103 with increased canonical Treg markers (FoxP3, HELIOS, CD25, CD39, CTLA-4, CCR4, and IL-10) and reduced pro-inflammatory cytokines. A unique cytotoxic program was marked by elevated Granzyme-K (GzmK) and Thrombospondin-4 (Tsp-4), a thrombospondin family extracellular matrix glycoprotein upregulated in activated CD8+ T-cells. Cytolysis was primarily mediated by Perforin (Prf) and multiple Granzymes packaged into Tsp-4&#x207a; supramolecular attack particles (SMAPs), with GzmK contributing to both cytotoxic and suppressive functions. After anti-CD19scFv CAR (CAR19) transduction, CAR19 + CD8-iTreg showed superior in vivo anti-tumor efficacy compared with CAR19-CTLs, significantly reducing tumor burden and prolonging survival in a CD19 + Nalm-6 human leukemia xenograft model while maintaining low pro-inflammatory cytokine production. In a xenogeneic graft-versus-host disease (GVHD) model with residual human leukemia, CAR19&#x207a; CD8-iTreg inhibited GVHD lethality and controlled tumor growth without increasing systemic inflammation. Together, these findings support CD8-iTreg-based CAR therapies as a strategy to retain potent anti-leukemic activity while limiting inflammatory toxicities of conventional CAR T-cells, properties particularly beneficial in treating auto- and allo-immune diseases. CD8-iTreg drive parallel tumoricidal and immunoregulatory functions mediated by releasing Tsp-4 + SMAPs containing granzyme K.

Cochlear implantation (CI) has emerged as a potential strategy for auditory rehabilitation in patients undergoing vestibular schwannoma (VS) surgery. However, outcomes remain variable, particularly in patients with neurofibromatosis type 2 (NF2). To evaluate auditory and patient-reported outcomes following simultaneous CI during VS resection. This prospective consecutive case series included six adult patients undergoing simultaneous translabyrinthine VS resection and CI. Three patients had sporadic VS and three had NF2-associated bilateral tumors. Audiological outcomes were assessed preoperatively (T0) and one year after CI activation (T1), including pure-tone average (PTA) and speech perception. Patient-reported outcomes included three questionaries. Auditory perception was achieved in all patients. Postoperative PTA was similar between sporadic and NF2 patients (29.5&#x2009;dB vs. 31.75&#x2009;dB). However, speech perception outcomes differed substantially. Sporadic VS patients achieved high performance (monosyllabic 90%, disyllabic 96%, sentence recognition (SR) 100%), whereas NF2 patients showed lower scores (monosyllabic 40.35%, disyllabic 38%, SR 35%). Tinnitus improved in all patients, with complete remission in 66.7% of cases. Simultaneous CI during VS surgery is a feasible rehabilitation strategy in selected patients. Speech perception differed between sporadic and NF2 cases despite similar hearing thresholds, highlighting the importance of cochlear nerve integrity, tumor burden and duration of deafness.

Emerging evidence links hearing loss to increased fall risk in older adults, yet mechanisms remain unclear. This study examined the impact of simulated hearing loss on reactive balance control in healthy young and older adults. Participants performed dual-task scenario involving auditory sentence repetition and unexpected surface translations. Hearing loss was simulated using Adobe Audition, applying frequency-specific attenuation based on moderate sensorineural hearing loss levels from the literature. Speech-in-noise performance significantly declined under simulated hearing loss for both age groups (p < .001), confirming effective simulation, but not significant differences were found between age groups. Maximum COP-COM displacement (the scalar distance between Center of Pressure and Center of Mass) increased with perturbation intensity (Level 1 vs. Level 2: p < .001), but no significant differences were found between hearing conditions or age groups (p > .05), suggesting compensatory step execution was preserved. Reaction time analysis revealed a significant interaction between hearing condition and perturbation level (p < .01). Older adults showed a smaller reduction in reaction time between Level 1 and Level 2 perturbations under simulated hearing loss (Level 1: 283&#xb1;37 ms; Level 2: 224&#xb1;20 ms) compared to normal hearing (Level 1: 298&#xb1;42 ms; Level 2: 230&#xb1;20 ms), indicating impaired ability to modulate initiation of reactive balance strategies during loss of balance with auditory challenge. These findings suggest that while older adults maintain compensatory stepping ability, simulated hearing loss may hinder their ability to adapt reaction timing during challenging dual-task scenarios. This impairment may contribute to increased fall risk in older adults with hearing loss. Further research is needed to explore underlying cognitive mechanisms and develop targeted interventions.

This study developed and validated monolingual and bilingual sentence-bidirectional encoder representations from transformers (SBERT) models for detecting cancer recurrence within Thai-English electronic medical records (EMRs) from Thai cancer hospitals. A multicentre dataset of 32&#x2009;436 documents from 1250 patients was used for model development. External validation involved an independent dataset of 9244 documents from 384 patients across two Thai cancer hospitals. Performance was benchmarked against a fine-tuned PubMedBERT (MetBERT). The development dataset included breast (43.9%), colorectal (12.1%), cervical (28.0%) and head and neck (16.0%) cancers. MetBERT achieved the highest area under the precision-recall curve (AUPRC) for locoregional versus no recurrence (11.1%) and locoregional versus distant recurrence (91.7%), while monolingual-SBERT excelled at distant versus no recurrence (32.0%). External validation demonstrated MetBERT superiority for locoregional versus no recurrence (9.30%-21.50%). For distant versus no recurrence, bilingual-SBERT performed best with AUPRC 17.55%-24.39%. While MetBERT led in distinguishing locoregional versus distant recurrence (88.30%-94.70%), bilingual-SBERT demonstrated robust external validation performance (AUPRC 85.25%-91.80%). Low AUPRC values (9%-32%) reflect the extreme class imbalance in real-world data (~1% recurrence prevalence). Despite this, fine-tuned MetBERT achieved highest performance, while bilingual-SBERT demonstrated superior robustness during external validation. This validates sentence embedding models for handling mixed Thai-English medical records in multilingual clinical environments. Sentence embedding frameworks provide a practical, generalisable solution for detecting cancer recurrence within multilingual EMRs. Despite text-length constraints, these models are suitable for clinical integration as a screening tool for cancer registry workflows.

Large language models (LLMs) are increasingly explored for qualitative analysis, but the effect of workflow design on thematic fidelity remains unclear. This study evaluated a structured human-AI collaboration framework using Claude Opus 4.6 to analyze 16 interview transcripts from patients with chronic obstructive pulmonary disease participating in a pulmonary telerehabilitation program. The workflow included code extraction, code combination, and theme generation, and was tested using hierarchical and direct strategies. AI-generated themes were compared with human-derived themes using sentence-t5-xxl embeddings and cosine similarity, with theme alignment performed using Hungarian and greedy matching. Output volume varied substantially across strategies, ranging from 53 to 357 codes and 11 to 17 themes. Direct grouping (average cosine similarity 0.891) and L3 grouping (0.890) achieved the highest similarity to human-generated themes. These findings suggest that grouping-based workflows can preserve key information, reduce redundancy, and improve thematic generation in LLM-assisted qualitative analysis.

Fibrosing skin diseases are highly morbid conditions with diverse clinical and histopathologic features. Prior work, primarily in systemic sclerosis (SSc), has yielded mixed data regarding the immune drivers of fibrosis, as well as the identity and spatial localization of pro-fibrotic fibroblast cell subsets. Here, we focus on morphea and eosinophilic fasciitis (EF), which cause more acutely inflammatory skin fibrosis. Using multimodal single-nucleus and spatial transcriptomics, we find that effector CD8+ T cells are highly enriched in fibrotic skin. These cells are particularly abundant in inflammatory tissue domains bordering fibrotic stroma, which are marked by expression of interferon-&#x3b3; stimulated genes. Inflammatory domains feature a loss of local homeostatic fibroblast populations and replacement with ADAM12-expressing inflammatory fibroblasts and myofibroblasts, which co-localize closely with CD8+ T cells. All subtypes of morphea featured similar patterns of CD8+ T-cell-associated fibro-inflammatory zonation and fibroblast transformation, suggesting that shared mechanisms can drive fibrosis across stromal compartments of skin. We apply these findings to a large publicly available scleroderma dataset and find that similar processes occur in SSc. Mechanistically, ablation of CD8+ T cells in mice ameliorates bleomycin-driven inflammation and fibrosis, as does fibroblast-intrinsic abrogation of IFN-&#x3b3; signaling. These data establish CD8+ T cell-driven fibrogenesis as a key feature of fibrosing skin diseases and raise the prospect of targeting CD8+ T cells in autoimmune fibrosis more broadly. Spatial profiling of morphea-spectrum diseases reveals CD8 T cells as key drivers of fibrosis through fibroblast IFN-&#x3b3; signaling.