Vallamchetla, Sai Krishna

Abstract Objective Evaluate how RAG architecture, including corpus structure, retrieval strategy, and pipeline complexity, affects LLM-based medical problem solving and knowledge retrieval in sleep medicine. Materials and Methods We benchmarked four open-source LLMs (Llama-3-8B, Llama -3 -70B, Qwen 2.5-14B, and Qwen 2.5-235B) using a knowledge base of five sleep medicine textbooks. We compared performance across three dimensions: corpus structure (raw text vs table-of-contents aligned.), retrieval strategy (dense embedding vs hybrid sparse-dense), and pipeline complexity (baseline vs augmented). Evaluation metrics included board-style multiple choice question (MCQ) accuracy and clinical case vignette diagnostic ranking. Results RAG improved MCQ accuracy for all models. Llama-8B saw the largest gain of 10.6% (61.8% to 72.4%), while Qwen-235B reached 87.3%. In clinical cases, Llama-8B accuracy dropped by 7.1% when using raw text and dense retrieval due to context noise. This was corrected by using structured hybrid configurations. Hybrid retrieval consistently outperformed dense-only methods. Overall, structured corpora improved primary diagnosis accuracy by 6.1% on average, with Qwen-235B reaching a peak 10.2% increase. Discussion RAG effectiveness depends on the balance between model size and data structure. Large models handle uncurated text well, but smaller models are easily distracted by irrelevant data. Hybrid retrieval is necessary to maintain precision with specialized medical terms. A structured corpus paired with a baseline hybrid pipeline offers the best stability and speed for clinical use. Conclusion Rigorous data curation and hybrid retrieval are as essential as model scale for deploying safe, guideline-compliant AI in sleep medicine.

Objectives: This study aims to identify retinal biomarkers of NOTCH3 mutation carriers using optical coherence tomography (OCT) data from the UK Biobank. Methods: We conducted a cross-sectional, matched case-control study of individuals with or without NOTCH3 mutation from the UK Biobank between 2006 and 2010. All participants had macular OCT scan and cognitive assessment. Cases were identified based on pathogenic/likely pathogenic NOTCH3 mutations and 1:1 matched with controls based on age, sex, hypertension, diabetes mellitus, and cigarette smoking status. Baseline characteristics and cognitive test scores were compared using χ 2 or Mann-Whitney U test appropriately. Macular thickness at central, inner, and outer subfields and at each retinal layer was compared using Wilcoxon signed-rank test. Results: Our analysis included 134 participants (67 cases, 67 controls) with a mean age of 54 ± 9 years. NOCTH3 mutation carriers performed worse in prospective memory ( P =0.043), verbal and numerical reasoning ( P =0.178), visual memory ( P =0.227), and processing speed ( P =0.101) than matched controls. There were no differences in visual acuity between the 2 groups. NOTCH3 mutation carriers had significantly thinner macular inner subfield at the superior ( P =0.006), temporal ( P =0.008), and nasal ( P =0.034) quadrants, and significantly thinner macular RNFL ( P =0.008) compared with age-, sex-, and vascular risk factor-matched controls. Conclusion: The presence of NOTCH3 mutation is associated with reduced thickness in the inner macular subfield and macular RFNL. These retinal changes may reflect early pericyte dysfunction and microvascular ischemia. Longitudinal studies are needed to assess the temporal relationship between these retinal changes, cerebrovascular disease progression, and clinical severity of disease.

Background: The management of asymptomatic high-grade carotid stenosis remains challenging, particularly in identifying individuals at highest risk for hypoperfusion-related stroke. Impaired cerebrovascular reactivity (CVR) has emerged as a potential predictor of future cerebrovascular events. Objective: To perform a systematic review and meta-analysis to quantify the association between impaired CVR and the risk of ipsilateral ischemic stroke or TIA in patients with asymptomatic high-grade carotid stenosis/occlusion. Methods: A systematic literature search of PubMed, Embase and Scopus was conducted from inception till July 2025. Eligible studies included asymptomatic patients with ≥70% carotid stenosis or occlusion who underwent CVR assessment and reported ipsilateral ischemic stroke or TIA outcomes. Pooled relative risks (RR) with 95% confidence intervals (CIs) were calculated using a random-effects model. Forest plots were generated, and heterogeneity was assessed using the chi-square test and the I 2 statistic. Results: Six prospective observational studies including a total of 361 patients with asymptomatic high-grade carotid stenosis/occlusion (mean age 70.4 years; 73.6% male; 71.2% with hypertension; 28.6% with diabetes; 29.6% current or former smokers) were included in the analysis. Of these, 120 patients had impaired CVR and 241 had preserved CVR. Transcranial Doppler (TCD) was the most frequently used modality to assess CVR, with acetazolamide or CO 2 as the vasodilatory stimulus. Following CVR assessment, patients were monitored for a weighted mean follow-up of 26.2 months (range: 21.5 to 42.5 months). The incidence of ipsilateral ischemic stroke or TIA was 14.2% in patients with impaired CVR versus 2.1% in those with preserved CVR (RR 3.98; 95% CI 1.62–9.76; P = 0.003; Figure 1). When limited to ipsilateral ischemic stroke alone, the incidence was 21.4% in the impaired CVR group compared to 4.3% in the preserved CVR group (RR 3.35; 95% CI 1.54–7.25; P = 0.002; Figure 2). Conclusions: Impaired CVR is associated with a 3- to 4-fold increased risk of ipsilateral ischemic stroke or TIA in patients with high-grade carotid stenosis/occlusion. CVR assessment may help stratify risk and guide individualized management in this population.

Background: Previous studies have reported the mean difference in homocysteine level in patients with and without silent brain infarction. However, evidence regarding the rate of silent brain infarcts in the presence versus absence of hyperhomocysteinemia remains limited. Objective: We aim to perform a systematic review and meta-analysis to compare the rate of silent brain infarcts in patients with hyperhomocysteinemia compared with patients with normal or low homocysteine levels. Methods: A systematic search of PubMed, Embase, Web of Science, Scopus were searched from inception to June 18, 2025. Eligible studies compared the rate of image-confirmed silent brain infarcts in patients with hyperhomocysteinemia versus those with normal homocysteine levels. The primary outcome was the rate of silent brain infarcts; the secondary outcome was cognition assessed by the Mini-Mental State Examination (MMSE). Pooled odds ratios (OR) with 95% confidence intervals (CI) were calculated using a random-effects model, and heterogeneity was assessed using I 2 statistics. Subgroup analyses were performed according to increasing hyperhomocysteinemia cutoffs (>12 μmol/L and 9–12 μmol/L) as defined in the included the studies. Results: Four observational studies involving 2331 patients were included (mean age was 58.0±6.4 years, 45.8% women). Of these, 1087 had hyperhomocysteinemia and 1244 had normal levels. The rate of silent brain infarcts was 23.3% (253/1087) in patients with hyperhomocysteinemia versus 22.8% (284/1244) in controls (OR 1.82; 95% CI 1.11-3.00; P = 0.02, Figure 1A). The association was higher in studies using a > 12 µmol/L cutoff (OR 3.08; 95% CI 1.85-5.11; P < 0.0001, Figure 1B). Patients with hyperhomocystinemia had lower MMSE score compared with controls (mean difference of -1.55; 95% CI -3.60-0.50; P = 0.14, Figure 2). Conclusions: Hyperhomocysteinemia is associated with increased risk of silent brain infarcts and numerically lower cognitive performance. Findings suggest a possible dose-dependent relationship between homocysteine level and silent stroke risk. Prospective studies are needed to determine whether homocysteine is a modifable therapeutic tharget for reducing silent infarcts and cognitive decline.

Background: Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy (CADASIL) can be misdiagnosed as multiple sclerosis (MS). This study aims to contrast demographic, clinical, neuroimaging, and genetic features of CADASIL patients with vs. without prior MS misdiagnosis. Methods: We conducted a retrospective cohort study to compare genetically or histopathologically confirmed CADASIL patients with and without a prior MS misdiagnosis across Mayo Clinic. Patients were classified into high-, medium-, low-, or unknown-risk by location of NOTCH3 mutation. Disease severity was assessed using NOTCH3 Small-Vessel Disease (N3SVD) scale. Between-group comparisons used independent t-test for age, chi-square or Fisher’s exact test for categorical variables, and Mann–Whitney U test for quantitative variables. Within the misdiagnosed subgroup, neuroimaging changes between MS to CADASIL diagnosis were assessed using Wilcoxon signed-rank test for quantitative data and Stuart–Maxwell test for categorical data. Significance was set at p<0.05 (two-sided). Results: Among 107 patients (mean age 57.5±13.4 yrs) with CADASIL, 12 (11.2%) had a prior MS misdiagnosis (Table 1). Compared to patients without misdiagnosis (n=95), those misdiagnosed were younger, more frequently Asian, more likely to undergo lumbar puncture (83.3% vs 26.3%, p <0.001), and less likely to have had stroke (33.3% vs 65.3%, p =0.031). There were no significant differences in deep white matter hyperintensity scores or lacune counts. However, the misdiagnosed group had a higher median modified Rankin Scale ( p =0.008) and a different distribution of N3SVD stages ( p =0.017), with a higher proportion in stage 3A (33.3% vs 5.3%). There was no significant difference in genetic risk categories (Figure), but 5 of 12 misdiagnosed cases clustered in the medium risk EGFR 13, 13.5, and 14 domains. Among the misdiagnosed, 41.7% received MS-specific treatment during a median of 10.2 months (range 0.4–138.1) before diagnostic correction (Table 2). Over this period, there were no significant changes in DWMH score, lacune count, mRS, or N3SVD stage (all p >0.05). Conclusions: About one in nine CADASIL patients had a prior MS diagnosis. Atypical presentations in younger individuals without a clear stroke history may prompt consideration of MS over CADASIL, leading to diagnostic delays and unnecessary treatments. EGFR domain risk category did not correlate with misdiagnosis.

Introduction: Studies suggest that stroke and hypoperfusion from heart failure (HF) may contribute to cognitive decline. However, the relationship between the HF subtypes and cognition remains poorly described in real-world populations. Hypothesis: We hypothesized that HF subtypes are differentially associated with cognitive impairment and that lower left ventricular ejection fraction (LVEF) correlates with poorer cognitive performance. Methods: We conducted a cross-sectional study of patients aged ≥18 years seen at a Heath System in 2022 who underwent both echocardiogram and cognitive assessment within one year. HF was classified into four groups based on 2022 American Heart Association HF guideline: No heart failure, HFpEF (HF with preserved LVEF), HFmrEF (HF with mid-range LVEF), and HFrEF (HF with reduced LVEF). Cognitive impairment was categorized using the Montreal Cognitive Assessment (MOCA) as normal (≥26), mild (18–25), moderate (10–17), and severe (<10). Multivariable linear regression assessed the correlation between EF and MoCA, and logistic regression evaluated the association between HF subtypes and severity of cognitive loss, adjusting for age, sex, race, hypertension, diabetes, dyslipidemia, chronic kidney disease, stroke. Results: A total of 1,695 patients were included in the study (mean age, 75.3±14 years; 48.3% female). 82.0% had HFpEF; 7.1%, HFmrEF; 7.9%, HFrEF, and 3.0% no HF. LVEF was positively associated with MoCA score (β=0.028, p=0.032). Patients with LVEF ≤40% had lower MoCA scores (17.7 ± 6.2 vs. 19.1 ± 5.9, p=0.008) and higher adjusted odds of moderate/severe cognitive impairment (aOR 2.49, 95%CI 1.20-5.15, p=0.014), compared to those with LVEF >40%. Adjusted odds of moderate to severe cognitive impairment were 1.14 for HFpEF, 0.99 for HFmrEF, and 2.81 for HFrEF compared to patients with no heart failure (trend-P=0.031, Table). Conclusions: Lower LVEF was independently associated with worse cognitive performance. Patients with HFrEF had a 2.5-fold higher risk of moderate to severe cognitive impairment. This association was not observed in patients with HFpEF or HFmrEF, suggesting differential cognitive effects across HF subtypes. Further studies are warranted to confirm these findings and evaluate the impact of interventions targeting EF and HF subtypes on cognitive outcomes.