YAYosef Adiniaev

Yosef Adiniaev

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Yosef Adiniaev has a total of 28 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    False-Positive Calls of Reimplemented Published Analyses for Covert Command-Following on Electroencephalograms Recorded Without Study Commands

    by Alon Gorenshtein, Yosef Adiniaev, Yiftach Barash, Eyal Klang, Oved Daniel

    Objective: To measure how often published electroencephalographic analyses for covert command-following call patients positive without commands, and their detection of genuine responses. Methods: This cross-sectional study applied 11 reimplemented published analyses to electroencephalograms without commands: 974 clinical electroencephalography patients and, with rules fixed, 730 new patients; 38 with prolonged disorders of consciousness; and 544 comatose patients after cardiac arrest. Outcomes: false calls (share called positive) against a 5% benchmark, and genuine-response detection. Results: Three of 10 command-following analyses called significantly more than 5% of clinical electroencephalography patients positive, replicated in new patients: Goldfine 2011, a spectral test, 23.3% (95% CI, 20.2%-26.6%), above its stated 5%; du Bois 2026, a criterion for entry to brain-computer interface training, 22.2% (95% CI, 19.2%-25.5%); and Guger 2018, a P300 accuracy threshold, 13.8% (95% CI, 11.5%-16.3%), near its approximate 12% chance rate (24.5% over 2 runs). The other 7 called 0.0% to 6.9%. After cardiac arrest: 11.5% to 27.3%. All 3 detected genuine responses. Identified contributors: a variance ignoring within-command dependence (Goldfine 2011), selection over configurations and times (du Bois 2026) and a permissive threshold (Guger 2018). Goldfine 2011’s command-level correction, fixed before testing, called 0.8% of new patients and 2.4% after arrest, detecting 93.4% of stroke imagery-versus-rest sessions but few between imagery types. Interpretation: Reimplemented published analyses called up to 27% of patients positive without commands, more with repeated runs. A positive result needs its complete call rule’s false-call rate and detection at that rate; imagery-versus-rest detection may reflect task engagement, not imagery.

  • Reporting of Fairness and Equity in Studies of Regulator-cleared Commercial AI for Acute Stroke CT: A Systematic Review

    by Yosef Adiniaev, Mahmud Omar, Tohar M. Timor, Oved Daniel, Yonatan Adiniaev, Yiftach Barash, Olga R. Brook, Eyal Klang, Alon Gorenshtein

  • Open Access

    Safety-Filter Fallback on Consumer Health Questions During the Initial Claude Fable 5 Deployment: Observational Study

    by Yosef Adiniaev, Mahmud Omar, Yiftach Barash, Olga R Brook, Alon Gorenshtein, Eyal Klang

    Background: Independent audits of medical large language models have concentrated on models that answer unsafely rather than those that decline safe questions. In June 2026, Anthropic released Claude Fable 5 with a safeguard that detects requests related to cybersecurity, biology, or chemistry and reroutes them to a fallback model, Claude Opus 4.8. The developer states that this safeguard is deliberately conservative and affects fewer than 5% of sessions. Whether that figure holds for consumer health questions, many containing dense biomedical vocabulary, was unknown. Objective: This study measured how often Claude Fable 5 routed consumer health questions to its fallback model and whether routing varied by clinical domain and framing. Methods: In this observational, point-in-time audit, the first 500 unique prompts in alphabetical order from the HealthSearchQA benchmark (3173 prompts) were each entered once into Claude Fable 5 through the web interface, in a new session with no system prompt and default settings, from June 9 to 12, 2026. Two reviewers independently coded each response as routed to fallback upfront or not routed upfront, with a third adjudicating disagreements. Midgeneration truncation due to a safety interruption was recorded separately. Prompts were labeled by clinical domain, question type, and sensitivity. Rates were reported with Wilson 95% CIs and compared using chi-square tests. Gemini 2.5 Flash served as an independent comparator. Results: Reviewers agreed on 474 of 500 responses (94.8%; Cohen κ=0.90; 95% CI 0.86-0.94). Fable 5 routed 243 (48.6%; 95% CI 44.2%-53.0%) prompts to the fallback. Of the 257 not routed upfront, 236 (91.8%; 95% CI 87.8%-94.6%) were fully answered and 21 (8.2%; 95% CI 5.4%-12.2%) were truncated midgeneration. Fallback routing varied by clinical domain ( χ 2 10 =59.5; P <.001), from 88.9% (32/36; 95% CI 74.7%-95.6%) for oncology and 79.3% (23/29; 95% CI 61.6%-90.2%) for reproductive and obstetric prompts to 26.3% (5/19; 95% CI 11.8%-48.8%) for mental and behavioral health. It also varied by question type ( χ 2 3 =61.9; P <.001): 63.1% (101/160; 95% CI 55.4%-70.2%) for prognosis or severity, 49.6% (137/276; 95% CI 43.8%-55.5%) for definition, and 1.8% (1/55; 95% CI 0.3%-9.6%) for diagnosis or treatment. Gemini answered 89.7% (218/243; 95% CI 85.3%-92.9%) of routed prompts and declined 19% (95/500; 95% CI 15.8%-22.7%) overall. Standardization to the full 3173-prompt question-type and domain compositions yielded fallback rates of 48.8% and 47.5%, respectively. Conclusions: Claude Fable 5 routed nearly half of these benchmark consumer health prompts away from the primary model, and routing was associated with clinical domain, question category, and disease vocabulary. Because benignness was not independently adjudicated, wording covaried with clinical content, and the fallback model’s subsequent answer was not recorded, these findings describe routing and truncation rather than user-facing refusal or a causal classifier feature. Fallback routing is a measurable safety property that audits should report alongside answer quality.

  • Open Access

    Limits of Trial-Adaptive Neural–Language Fusion Across Large Language Models in P300 Brain–Computer Interfaces

    by Alon Gorenshtein, Mahmud Omar, Eric Jia, Yosef Adiniaev, Oved Daniel, Jonathan Kruskal, Muneeb Ahmed, Olga Brook, Eyal Klang, Yiftach Barash

    Objective: Published P300-speller fusion schemes fix prior trust regardless of trial reliability; we tested whether a reliability estimate improves on it. Methods: We reanalyzed 3,373 archived P300-speller selections from 47 people with ALS (BigP3BCI). A fair, matched-search-space comparison, tuning both a fixed weight and an adaptive policy out-of-fold, was evaluated across 22 evaluable language-model priors up to 46.7B parameters. Two representative priors, GPT-2 and a classical 5-gram, additionally received detailed naive and mechanistic analyses. Results: No prior’s 95% CI favored adaptive fusion under the fair comparison, despite unexploited oracle headroom at every scale. Under GPT-2, the naive comparison was significantly worse for adaptive fusion; both anchors converged to a degenerate or near-degenerate fair-comparison solution. For the representative anchors, three further controllers failed to convert that headroom into benefit; the fixed-fused posterior’s output probability outperformed the best controller for flagging errors (2.8- to 3.8-fold enrichment). Conclusion: A tuned fixed weight is a difficult-to-beat default across the tested scale range; reliability estimation gave no deployable adaptive advantage. Significance: Adaptive weighting should be validated against a fairly tuned baseline across model families and scales; in this dataset, the fused output’s confidence identified high-risk selections better than the tested purpose-built ranker.

  • Open Access

    Nighttime intramedullary nailing of femoral fractures is not associated with worse outcomes: a 13-year Level I trauma center cohort

    by Artsiom Abialevich, Vadim Benkovich, Vladislav Osinsky, Yosef Adiniaev, Ilan Tzaytlin, Yotam Heinemann, Asaf Acker

    Background: Whether nighttime orthopedic trauma surgery adversely affects outcomes remains controversial. Evidence specifically addressing intramedullary fixation across femoral fracture locations remains limited. This study evaluated whether nighttime surgery was associated with a higher rate of recorded postoperative complications than daytime surgery. Methods: This retrospective observational cohort included adult patients with nonperiprosthetic femoral fractures treated with intramedullary fixation between January 2010 and January 2023 at a single Level I trauma center. Of 311 screened records, 23 were excluded because the patient was younger than 18 years ( n = 17) or the fracture was periprosthetic ( n = 6), leaving 288 procedures (242 daytime and 46 nighttime). The primary outcome was any recorded postoperative complication during available follow-up. Secondary outcomes included radiographic abnormality, delayed union (> 24 weeks), nonunion, revision surgery, and mortality. Surgeon experience was categorized as senior surgeon or resident. Exploratory fracture-location subgroup analyses and multivariable logistic regression were performed. Results: Nighttime patients were younger (35.1 ± 19.1 vs. 46.6 ± 24.6 years, p < 0.001), more frequently male (82.6% vs. 64.0%, p = 0.014), more often injured by a high-energy mechanism (80.4% vs. 56.6%, p = 0.002), and underwent surgery sooner after trauma (4.0 [2.0–10.0] vs. 18.0 [7.0–48.0] hours, p < 0.001). Residents were the recorded primary surgeon in 54.3% of nighttime and 39.3% of daytime procedures ( p = 0.057). After adjustment for demographic, injury, fracture-location, time-to-surgery, and surgeon-experience variables, nighttime surgery was not associated with the primary outcome (adjusted odds ratio 0.45, 95% confidence interval 0.15–1.38; p = 0.163). Delayed union (28.3% vs. 19.4%, p = 0.233) and revision surgery (4.3% vs. 12.4%, p = 0.130) did not differ significantly. Conclusions: In this single-center cohort, nighttime intramedullary fixation of femoral fractures was not associated with higher observed rates of postoperative complications, impaired fracture healing, or revision surgery compared with daytime fixation. These findings suggest that operative timing alone may not be a major determinant of outcome within an organized Level I trauma system. Nevertheless, the small nighttime cohort, unequal group sizes, fracture heterogeneity, and limited statistical precision require cautious interpretation and preclude definitive conclusions regarding equivalence or subgroup-specific safety. Larger multicenter studies with standardized assessment of surgeon experience, staffing, fatigue, and fracture-specific outcomes are warranted.

  • Open Access

    Quantifying Large Language Model Influence in Brain Computer Interface Communication for Amyotrophic Lateral Sclerosis

    by Alon Gorenshtein, Mahmud Omar, Eric Jia, Yosef Adiniaev, Oved Daniel, Jonathan Kruskal, Muneeb Ahmed, Olga Brook, Yiftach Barash, Eyal Klang

    Large language models are integrated into brain-computer interfaces for communication, but accuracy does not show whether an emitted character depended on neural evidence or on the language-model prior. We retrospectively re-decoded 3,373 P300-speller selections from 47 people with amyotrophic lateral sclerosis, reconstructing a neural posterior and combining it with 25 language priors ranging from 5-grams to 46.7-billion-parameter models. Under held-out, per-source calibration and equal fusion weighting, the prior accounted for a participant-weighted mean 8.6% of posterior displacement (median across selections, 2.9%); the corresponding neural contribution fraction was 0.914 (95% CI, 0.896-0.934). In 4.4% of selections (95% CI, 3.5-5.3), the fused system emitted the intended character although neural evidence alone would not have selected it. Results were similar across 21 neural language models. Accuracy alone does not reveal how strongly a fused brain-computer interface decision depends on its language prior.

  • Open Access

    Levodopa Administration Timing During Hospitalization: Associations With Intensive Care Unit Exposure and Documented Access Type

    by Alon Gorenshtein, Mark Katson, Yosef Adiniaev, Eyal Klang, Oved Daniel

    Background: Levodopa is time-critical in hospitalized Parkinson disease. Whether dosing fidelity depends on care setting or documented access status is unclear. Objectives: To quantify levodopa dosing fidelity, test ICU exposure with clustering-aware methods, and test whether documented access type is associated with delayed or omitted dosing. Methods: Retrospective cohort study using MIMIC-IV (2011-2022). Adults with Parkinson disease and ≥1 scheduled levodopa dose contributed 1,665 admissions and 39,322 doses. ICU exposure was tested with a patient-clustered GEE model. Among ICU-exposed doses, access type (normal, tube feeding, parenteral nutrition, NPO) was modeled in one fully adjusted model and tested for specificity, restricted to the ICU, against an active-comparator medication (statins). Results: Of 39,322 doses, 79.8% were on time by the primary 60-minute definition; a symmetric ±15-minute definition classified 68.8% as mistimed. ICU exposure was not associated with delayed or omitted dosing after clustering (patient-clustered OR, 0.87; 95% CI, 0.74-1.01). Among ICU-exposed doses, NPO was associated with delayed or omitted dosing (adjusted OR, 1.89; 95% CI, 1.36-2.62) and tube feeding with lower odds (adjusted OR, 0.62; 95% CI, 0.42-0.92; P < .001). The comparator medication showed a directionally consistent but inconclusive interaction (OR, 1.27-1.28; 92 patients). A route-order association was not observed among immediate-release formulations (OR, 0.72; 4 patients). Conclusions: ICU admission alone was not associated with dosing unreliability after clustering. Among ICU-exposed doses, access type, not a single pooled category, was associated with dosing reliability; a comparator-medication check, valid only in the ICU, was directionally consistent but inconclusive.

  • Open Access

    Calibration-derived decoder discriminability is associated with online P300-speller accuracy, but the fitted mapping does not transport across cohorts

    by Alon Gorenshtein, Yosef Adiniaev, Tom Liba, Yiftach Barash, Eyal Klang, Oved Daniel

    Objective: A calibration-derived score has been related to online P300-speller accuracy in the same session, always within the cohort measured. We evaluated whether a fitted mapping from that score to expected accuracy transports to withheld cohorts. Approach: Retrospective secondary analysis of BigP3BCI 1.0.0, the largest published evaluation of this relationship to date: 18 of 20 source studies yielded eligible online outcomes, contributing 271 participants, 739 session-condition records, and 19,611 character selections. Four cohorts document an amyotrophic lateral sclerosis (ALS) population. The predictor was calibration-derived decoder discriminability: cross-validated discriminability of a classifier fitted only to calibration epochs. One source study at a time was withheld from development. Between-cohort variation was summarised by the random-effects standard deviation tau, with participant-clustered standard errors. Main Results: The association was positive in all 18 cohorts but varied widely in magnitude and precision (Pearson r 0.190 to 0.928; participant-level r = 0.714, p < 0.001). Pooled estimation error was 0.098 (95% CI 0.091 to 0.107) against a benchmark of 0.146. Calibration did not transport: the intercept had tau 0.87, with a 95% interval for an unrepresented cohort of -1.97 to 1.85 on the log-odds scale, and the slope varied more than tenfold across cohorts (tau 0.43, unrepresented-cohort interval 0.111 to 2.005). A protocol proxy for the stopping rule, median time per selection, reduced the between-cohort slope variance by 66%. The four ALS cohorts alone gave an uncorrected between-cohort slope spread of 0.223, against 0.560 overall. Significance: The score carries a reproducible signal about accuracy in the corresponding session, but the mapping between them is cohort-specific: usable for ranking sessions within a setting, not for reporting expected accuracy elsewhere without local recalibration. Few-cohort evaluation, as the ALS subgroup illustrates, understates how much performance varies elsewhere.

  • ChatGPT and Other Large Language Models in Inflammatory Arthritis: A Systematic Review Across Clinical Tasks

    by Yosef Adiniaev, Mahmud Omar, Tohar M. Timor, Yiftach Barash, Olga R. Brook, Mohammad E. Naffaa, Alon Gorenshtein, Eyal Klang

    Objective: Large language models (LLMs) are increasingly evaluated for rheumatology tasks, but their performance in inflammatory arthritis (IA) remains unclear. We systematically reviewed LLM performance across clinical tasks in IA. Methods: We conducted a systematic review (PROSPERO: CRD420261359100), searching PubMed, Scopus, and PubMed Central from January 2022 to April 2026 for studies evaluating LLM performance on clinical tasks in IA. Two reviewers (YA, AG) screened 113 records. Results: Eighteen studies covered rheumatoid arthritis (n = 3), axial spondyloarthritis (n = 7), psoriatic arthritis (n = 2), gout (n = 1), juvenile idiopathic arthritis (n = 1), and multiple diseases (n = 4). Most diseases and tasks were represented by only 1 to a few studies, and the evidence base remains early stage and uneven across conditions. Over 20 distinct LLMs were evaluated, including ChatGPT-3.5 to ChatGPT-4o, Gemini 2.0, DeepSeek-R1/V3, Claude, and Perplexity; ChatGPT/GPT variants were the most frequently tested models (16/18 studies), so the current evidence base is predominantly ChatGPT/GPT-based. Findings spanned patient education (n = 11), guideline adherence (n = 6), clinical reasoning (n = 3), and other applications (n = 1). All readability assessments exceeded the recommended thresholds. Guideline concordance ranged from 48% to 96%. Accuracy was lower for case-based clinical scenarios (4.24/6) than for frequently asked questions and guideline-based questions (5.32-5.36/6; P = 0.04). When compared with real clinical data, agreement was poor (Cohen and Fleiss κ ≈ 0). Conclusion: LLMs may support patient education, factual medication queries, and structured guideline questions when used under clinician review, but should not be used for case-based reasoning, treatment selection, or autonomous clinical decisions. None of the 18 included studies evaluated retrieval-augmented or agent-based systems, and none prospectively validated LLMs in clinical workflows. Safe integration in rheumatology will require purpose-built, knowledge-grounded systems and prospective evaluation before routine clinical use.

  • Open Access

    First-Line Opioids and Short-Term All-Cause Emergency Department Return After Headache Visits: A Two-Center Comparative Cohort Study

    by Alon Gorenshtein, Yosef Adiniaev, Tom Liba, Eyal Klang, Oved Daniel

    Objective: To compare first-line emergency department (ED) treatment classes for acute headache on short-term all-cause ED return and index admission across two independent health systems. Background: ED trials of acute headache treatment are judged on in-ED pain relief, a documented endpoint that is recorded incompletely and shifts with the scoring rule, and is a weak surrogate for what happens after discharge. All-cause ED return after an index headache visit (any subsequent ED encounter within the window) has not been used to compare first-line treatments at scale, and society guidance favors dopamine-receptor antagonists while recommending against routine opioids. Methods: Retrospective two-center cohort of adults treated for headache in the ED, using MIMIC-IV-ED (Beth Israel Deaconess Medical Center, 2011-2019) and MC-MED (Stanford, 2020-2022). The first-line class was the earliest qualifying acute agent. The primary contrast was opioids versus dopamine-receptor antagonists (the guideline-preferred class). Outcomes were 72-hour and 7-day all-cause ED return (among discharged patients; any subsequent ED encounter within the window) and index hospital admission. Confounding by indication was addressed with propensity overlap weighting; associations are reported as adjusted risk ratios (RRs) with bootstrap 95% CIs and E-values. Estimates were pooled with a site term and examined per site. Results: Among 13,285 treated adults (10,799 MIMIC-IV-ED; 2,486 MC-MED), opioid recipients were older and higher-acuity than dopamine-antagonist recipients (index admission 38.1% vs 16.4%). In the MIMIC-IV-ED discharged primary-contrast population, overlap weighting reduced the maximum standardized mean difference from 0.35 to 0.002; pooled and site-specific balance diagnostics are provided in the Supplement. First-line opioids remained associated with a higher 72-hour all-cause ED return (6.8% vs 3.8%; adjusted RR 1.79; 95% CI 1.31 to 2.33), 7-day return (10.7% vs 6.6%; RR 1.62; 95% CI 1.28 to 1.98), and index admission (RR 2.32; 95% CI 2.11 to 2.58, consistent with strong residual severity differences in patients selected for opioids). The direction of association was concordant across both health systems, although MC-MED return estimates were imprecise given the smaller opioid-treated discharged sample. In MIMIC-IV-ED, the cumulative all-cause return incidence by treatment class separated by day 3 and persisted through 30 days. The direction was consistent, though attenuated and no longer statistically significant, when the outcome was restricted to a headache-specific return (72-hour RR 1.31; 95% CI 0.91 to 1.88); the direction persisted for the composite of admission or 72-hour return, which does not condition on discharge but is influenced by the more confounded admission component (RR 2.16; 95% CI 1.98 to 2.39). Conclusion: Across two health systems, first-line opioid treatment for ED headache was associated with higher all-cause short-term ED return among discharged patients and higher index admission than dopamine antagonists. These observational associations reflect downstream all-cause ED utilization after an index headache visit rather than confirmed headache recurrence or treatment failure; they are consistent with guideline-concordant, opioid-sparing first-line treatment and warrant prospective confirmation. Plain Language Summary: Emergency departments treat headaches with several different medicines, but the usual way of judging which works, the pain score recorded during the visit, is often missing or inconsistent. Using two large hospital systems and a clearer outcome, whether patients came back to the emergency department for any reason, we found that patients first treated with opioids returned within 72 hours about 1.8 times as often as those given the guideline-preferred dopamine-blocking medicines and were admitted more than twice as often. These patterns pointed the same direction in both hospital systems after adjustment for the measured differences available in both databases. Because this was an observational comparison and returns were counted for any reason, the findings are consistent with using guideline-preferred non-opioid medicines first, rather than proof that opioids worsen headache. Graphical Standalone visual summary of the two-center comparison: among adults treated for headache in the emergency department, first-line opioids were associated with a higher all-cause ED return and higher index admission than guideline-preferred dopamine antagonists, whereas the headache-specific return was attenuated and crossed the line of no difference. Index admission is shown as a secondary acute-care outcome and should be interpreted cautiously because it is highly susceptible to confounding by baseline severity.