PaperlayerView exact revision on arXiv

Source-linked AI summary

GUIDE: Computationally Augmented Consensus for Trustworthy Clinical Practice Guidelines

Yuanfan Lin, Jie Di, Xujin He, Zhenzhe Lin, Jialing Chen, Ruiyang Li, Changyu Zheng, Yuanrui Shang, Weiqin Peng, Jingxi Qin, Xiong Wang, Dongxin Ye, Yining Huang, Guoqiang Li, Andi Xu, Sijia Yu, Ruixin Wang, Yahan Yang, Xiaohang Wu, Duoru Lin, Jingjing Chen, Zhenzhen Liu, Wenben Chen, Haotian Lin

arXiv:2610.16108v1cs.AIcs.CLcs.HCcs.MA

TL;DR

Rapid evidence growth and judgment-intensive consensus make timely guideline updates difficult. GUIDE combines AI agents with expert oversight, and the paper reports robust performance across benchmark, ablation, temporal hold-out, and prospective evaluations.

  • Problem

    Rapid biomedical evidence growth and the need for expert deliberation over uncertain evidence challenge timely clinical guideline generation and updating.

    Source: Introduction · p. 3

  • Method

    GUIDE uses a stepwise human-AI workflow in which LLM agents process evidence and experts provide judgment, oversight, and final deliberation.

    Source: Development of the GUIDE for clinical… · p. 5, Discussion · p. 16

  • Results

    GUIDE was a robust collaborative partner across benchmark evaluation, ablation analysis, temporal hold-out testing, and prospective real-world validation, reducing experts’ manual workload.

    Source: Discussion · p. 17

  • Takeaways & Limitations

    GUIDE redirects expert effort from repetitive tasks toward judgment, arbitration, and consensus-building in guideline development.

    Source: Discussion · pp. 17–18

    The benchmark’s scale and diversity remain modest, and more complex or controversial clinical questions must be tested to confirm generalizability.

    Source: Discussion · p. 18

Abstract

from arXiv · show

Maintaining up-to-date clinical practice guidelines (CPGs) is essential for evidence-based medicine, but rapid evidence growth and complex expert consensus formation limit timely updates. We developed GUIDE, a multi-agent copilot system for human-AI collaborative guideline generation and updating. GUIDE delegates labor-intensive tasks to specialized AI agents while preserving expert judgment and final arbitration. In 12 NICE ophthalmology datasets, GUIDE achieved 88.85% literature recall, 86.06% alignment in certainty-of-evidence ratings, and 91.67% recommendation agreement, requiring 162.17 minutes per dataset on average (28.68% human intervention). Performance remained stable across 20 additional datasets from cardiology, endocrinology, gastroenterology, and respiratory medicine. Ablation analysis showed that removing expert intervention reduced performance in judgment-intensive steps such as Risk-of-Bias assessment. In six post-cutoff datasets, GUIDE maintained robust end-to-end performance on unseen evidence. Prospective real-world validation showed that GUIDE reduced expert workload and produced more clinically acceptable recommendations than the original decade-old guideline (mean score, 3.73 vs. 1.53; agreement, 66.7% vs. 13.3%), with outcomes comparable to human-only updating (P = 0.588) but substantially less time and effort. GUIDE offers a practical framework for timely, trustworthy, expert-centric guideline development and validates computationally augmented consensus.

Introduction

Clinical guidelines struggle to keep pace with rapidly growing evidence and complex expert deliberation. GUIDE pairs AI processing with expert oversight and is evaluated across benchmark, specialty, temporal hold-out, and prospective settings.

Introduction

  • Conventional guideline updates lag behind rapidly growing biomedical evidence, with typical cycles exceeding two years and more than one million citations added annually.The scale and complexity of new evidence strain efficient synthesis.Source: Introduction · p. 3
  • Guideline development requires experts to interpret uncertain evidence, assess methodological limits, weigh benefits and harms, and reconcile clinical, economic, and societal considerations.Many steps depend on judgment and deliberation rather than deterministic rules.Source: Introduction · p. 3
  • Prior LLM systems support screening, extraction, and recommendation generation, but transparent evidence scrutiny, expert judgment, and accountable consensus under uncertainty remain inadequately addressed.The paper notes that LLMs can struggle with implicit inference, quantitative reasoning, and methodological nuance.Source: Introduction · pp. 3–4
  • GUIDE allocates high-volume, rule-based guideline tasks to AI agents while experts oversee the process, address flagged issues, and fine-tune outputs.The evaluation includes 12 NICE ophthalmology datasets, 20 datasets across other specialties, six post-cutoff datasets, and prospective updating of an intraocular-lens guideline.Source: Introduction · pp. 4–5

Results

GUIDE combines AI automation with expert oversight across guideline development, then evaluates benchmark performance, cross-specialty generalizability, and a prospective cataract update. Results show strong recommendation agreement and higher expert-rated acceptability than the original guideline, while ablation results underscore the value of human judgment in demanding stages.

Development of the GUIDE for clinical guideline generation

  • GUIDE delegates standardized tasks to agents but reserves PICO construction, evidence-quality interpretation, and final recommendations for experts, with proactive and trigger-based oversight.The workflow spans question decomposition through recommendation generation, with experts reviewing flagged discrepancies and validating final outputs.Source: Development of the GUIDE for clinical… · pp. 5–7

0.83 (ranging from 0.80 to 0.86). Across 12 benchmark tasks, the retrieval module identified

  • The pipeline achieved 88.85% recall at full-text screening, 86.06% final certainty-grading performance, and 91.67% consistency with original-guideline recommendations; expert oversight and refinement comprised 28.68% of total time.Source: §0.83 · pp. 8–9
  • Extraction averaged 28.20 minutes per dataset at 98.02% accuracy, but its 1.98% errors were numerical omissions that reviewer checks did not flag.Numerical values made up 88.9% of the items requiring expert verification.Source: §0.83 · pp. 8–9

Ablation analysis of expert and multi-agent contributions to performance

  • Without reviewer and expert input, final screening recall fell from 88.85% to 79.80%, precision from 84.60% to 66.09%, and risk-of-bias agreement from 82.80% to 69.54%.The ablation also found lower data-extraction accuracy and lower consistency in certainty ratings.Source: Ablation analysis of expert and… · pp. 9–10

Evaluation of GUIDE generalizability across clinical specialties

  • Across 20 NICE-derived datasets in four specialties, GUIDE achieved 88.83% full-text screening recall and 90.00% expert agreement on draft recommendations.The specialties were cardiology, endocrinology, gastroenterology, and respiratory medicine.Source: Evaluation of GUIDE generalizability… · pp. 10–11
  • On six post-cutoff datasets, final screening recall was 87.75%, and experts found 100.00% directional agreement with the original recommendations.Minor shifts in the evidence pool limited direct comparison of some intermediate metrics with the original NICE reference.Source: End-to-end evaluation of GUIDE… · pp. 11–12

Prospective case study of GUIDE for cataract guideline updating

  • Experts manually added resource-utilization and cost outcomes omitted from AI-generated PICO criteria; the updated evidence body comprised 27 eligible studies and supported shifting multifocal and EDOF IOL guidance from “do not offer” to “consider”.Source: Prospective case study of GUIDE for… · pp. 12–14
  • 3.73±0.96 mean score for GUIDE-assist versus 1.53±1.06 for the original 2017 NICE statement.Agreement was 66.7% versus 13.3%; GUIDE-assist and human-only versions did not differ significantly (adjusted P=0.588). The assisted update took 192 minutes, including 65 minutes of expert involvement, versus 23 days for human-only updating.Source: Prospective case study of GUIDE for… · pp. 14–16

Discussion

GUIDE combines LLM agents with expert oversight to support guideline development, shifting repetitive work to AI while retaining expert judgment for interpretation and consensus. Prospective validation and other evaluations support its efficiency and robustness, although its scope and several workflow capabilities remain limited.

Discussion

  • GUIDE’s workflow assigns repetitive, high-volume processing to AI and reserves interpretive, value-sensitive work for experts, including risk-of-bias assessment and recommendation formulation.Experts can correct anomalies, guide agents, and revise outputs; the system is framed as computationally augmented consensus rather than automation alone.Source: Discussion · pp. 16–17
  • By redistributing work, GUIDE reduces expert time on repetitive tasks and redirects attention toward judgment, arbitration, and consensus-building.The discussion emphasizes preserving expert attention for decisions that shape clinical validity.Source: Discussion · pp. 17–18
  • Across benchmark, ablation, temporal hold-out, and prospective evaluations, GUIDE operated as a robust collaborative partner and achieved comparable evidence synthesis and recommendation updating with greater efficiency than human-only updating.The discussion reports time savings in labor-intensive study selection and data extraction, while prospective validation reinforced the need for expert arbitration and contextual judgment.Source: Discussion · pp. 17–18
  • The benchmark’s scale and diversity remain modest, and testing on more complex or controversial questions is needed to confirm generalizability.The workflow also lacks automated extraction of primary studies nested in retrieved reviews, integrated economic evaluation, and modeling; expert panels still synthesize cost-effectiveness and resource-allocation evidence manually.Source: Discussion · pp. 18–19
  • Risk-of-bias assessment and final recommendation formulation remain sensitive to LLM reasoning nuances and can vary from human consensus; reviewer intervention still relies on static triggers.The authors identify broader multi-center trials as necessary to assess robustness and long-term clinical impact across organizational contexts.Source: Discussion · p. 19

Overall study design and datasets

The evaluation draws on NICE guideline materials to construct ophthalmology benchmarks, cross-specialty datasets, post-cutoff tests, and a prospective real-world updating case. The original materials for each dataset supply ground truth for modular evaluation.

Overall study design and datasets

  • 12 datasets from four ophthalmology guidelines were established for benchmark testing and ablation, with each dataset corresponding to the complete original materials for one review question.The selected NICE guidelines provide traceable search strategies, explicit eligibility criteria, and detailed evidence profiles.Source: Overall study design and datasets · p. 19
  • 20 additional datasets from six guidelines covered cardiology, endocrinology, gastroenterology, and respiratory medicine to assess cross-disciplinary generalizability.Source: Overall study design and datasets · p. 20
  • Six datasets based on four post-training-cutoff guidelines were used for end-to-end simulated prospective evaluations to test for data leakage.Ground truth included PICO elements, search strategies, study-selection records, extraction tables, evidence assessments, and final recommendations.Source: Overall study design and datasets · p. 20
  • A real-world prospective case used a removed recommendation from NICE’s 2017 cataract guideline to evaluate an actual guideline-updating scenario.Source: Overall study design and datasets · p. 20

Development of human-AI collaboration principles

A multidisciplinary expert panel used a three-round modified Delphi-like process to develop human-AI collaboration principles for GUIDE. Anonymized feedback and focused review supported refinement and resolution of contested items.

Development of human-AI collaboration principles

Engineering design of GUIDE framework

GUIDE uses specialized LLM agents, an AI reviewer and assistant, and medical experts across guideline development. Its workflow combines automated processing with expert validation, triggered review, and expert-led deliberation.

Engineering design of GUIDE framework

  • The multi-agent framework pairs task-specific LLM agents with an AI reviewer, an AI assistant, and medical experts who intervene when judgment is required.Standardized prompts define each agent’s role, instructions, and output format; templates were refined using five additional datasets excluded from evaluation.Source: Engineering design of GUIDE framework · p. 21
  • PICO construction and a draft review protocol precede retrieval, with experts validating the PICO and refining protocol parameters before searches begin.Experts can add contextual elements or broader considerations for complex questions.Source: Engineering design of GUIDE framework · p. 21
  • Two independent agents generate search strings and retrieve records; the reviewer flags substantial disagreement using a greater-than-30% record-count difference or Jaccard similarity below 0.5.Human experts can revise the strategy, and results from both retrieval rounds are merged to form the candidate set.Source: Engineering design of GUIDE framework · pp. 21–22
  • Independent agents screen titles and abstracts, then full texts, with discordant full-text decisions triggering expert review before included studies form the evidence body.GUIDE supports PDF acquisition, manual uploads, and OCR, and experts can inspect and revise screening decisions.Source: Engineering design of GUIDE framework · p. 22
  • Dual agents extract study and outcome data into evidence tables, while experts resolve discrepancies and judge whether studies are sufficiently homogeneous for quantitative synthesis.Where appropriate, the system selects a prespecified meta-analytic approach and uses R scripts to calculate pooled estimates.Source: Engineering design of GUIDE framework · pp. 22–23
  • Certainty assessments cover risk of bias, indirectness, imprecision, and inconsistency under GRADE, with discordant parallel-agent ratings referred for expert adjudication.The risk-of-bias tools are matched to study design, and domain assessments are synthesized into an overall certainty rating.Source: Engineering design of GUIDE framework · pp. 23–24
  • A Delphi-like AI committee integrates evidence and quality assessments into EtR materials and a preliminary recommendation, with experts monitoring deliberation and steering between rounds.Experts select from eight specialist roles or introduce new roles; the chair identifies convergence and divergence and continues deliberation until consensus criteria are met. Experts did not join final deliberation in retrospective experiments to avoid bias from prior knowledge of original recommendations.Source: Engineering design of GUIDE framework · p. 24

Implementation of modular test

The modular evaluation tested GUIDE components in isolation using gold-standard intermediate inputs, enabling direct comparison with reference materials. Evaluation covered clinical-question parsing, literature screening, data extraction, evidence quality assessment, and recommendation consistency.

  • Gold-standard intermediate results initialized each component in isolation, enabling direct comparison with reference materials and mitigating error propagation.This modular approach was used in the benchmark, ablation, and cross-specialty generalization tests.Source: Implementation of modular test · p. 24
  • Each benchmark and cross-specialty dataset was paired with a senior domain specialist with over ten years of relevant clinical experience.The specialist served as the human collaborative partner.Source: Implementation of modular test · p. 25
  • Literature retrieval used source-guideline PICO elements and eligibility criteria, with the retrieval period capped at the guideline’s publication date.Two-stage screening followed the source guideline’s search strategy, and screening was assessed by precision, recall, NPV, and specificity.Source: Implementation of modular test · p. 25
  • Extraction accuracy was checked against GRADE-table data, while risk of bias, inconsistency, indirectness, and imprecision informed overall evidence certainty.After Delphi-like deliberation, recommendations were classified as exactly consistent, partially consistent, or inconsistent with the original guidelines.Source: Implementation of modular test · p. 25

Implementation of end-to-end experiments using post-cutoff datasets

The end-to-end simulation used post-release guideline datasets to assess GUIDE on unseen evidence and reduce the risk that performance reflected training-data leakage. Unlike modular tests, each stage processed the preceding stage’s outputs.

  • Datasets from guidelines published after the release dates of GUIDE’s LLMs were used to assess end-to-end performance while addressing potential training-data leakage.The LLMs named were ChatGPT-4.1 and DeepSeek-R1.Source: Implementation of end-to-end… · p. 26
  • Unlike modular evaluation with gold-standard inputs, the simulation began with clinical questions and passed each stage’s generated outputs to its successor.PICO frameworks fed retrieval, retrieved literature fed two-stage screening, and later stages relied exclusively on studies selected earlier in the run.Source: Implementation of end-to-end… · p. 26
  • Each dataset was paired with a senior specialist with over ten years of clinical experience in the corresponding field.The specialist served as the human collaborative partner.Source: Implementation of end-to-end… · p. 26

Design and implementation of the prospective case study

The prospective case study compared GUIDE-assisted and human-only updating for a NICE cataract guideline question prioritized for updating. Independent cataract specialists conducted the GUIDE process, while blinded external physicians assessed the original and updated recommendations.

  • Two cataract specialists with over twenty years of clinical expertise conducted the prospective GUIDE update of the NICE 2017 cataract guideline.They independently assessed issues requiring human judgment and discussed disagreements to consensus; one operated GUIDE using only actions agreed by both.Source: Design and implementation of the… · p. 26
  • The case study addressed the question of the optimal strategy for simultaneous distance and near vision after cataract surgery.The question had been officially prioritized for updating on the NICE website.Source: Design and implementation of the… · p. 26
  • The parallel human-only process used an evidence review group and a decision committee to prepare evidence and formulate a draft recommendation.The committee used a modified three-round Delphi process and included clinical, provider, commissioner, and lay representation.Source: Design and implementation of the… · pp. 26–27
  • Fifteen external senior physicians, blinded to recommendation source, rated the original, GUIDE-assisted, and human-only statements in randomized order.Ratings used a five-point Likert scale and were classified as approval for scores of 4 or 5, or non-approval for scores of 1–3.Source: Design and implementation of the… · p. 27

Statistical analysis

Evaluation metrics were matched to task type, and results and recommendation approvals were analyzed using specified statistical procedures. Analyses used two-sided tests with a significance threshold of P < 0.05.

  • Clinical-question parsing used BERTScore, retrieval and screening used precision, recall, NPV, and specificity, and extraction and quality assessment used classification accuracy or deviation from reference values.Source: Statistical analysis · p. 27
  • Means were reported with 95% CIs from 1,000,000 bootstrap resamples, with paired t-tests or Wilcoxon signed-rank tests selected according to difference distributions.Source: Statistical analysis · p. 28
  • Recommendation approvals were reported as counts and proportions, compared overall with Cochran’s Q, and followed by Holm-adjusted exact McNemar tests when significant.All tests were two-sided, with P < 0.05 considered statistically significant.Source: Statistical analysis · p. 28

Data availability

The study used publicly available NICE guideline documents across ophthalmology, cardiology, endocrinology, gastroenterology, and other clinical areas.

  • Study materials included NICE guidelines for diabetic retinopathy, cataracts, age-related macular degeneration, and glaucoma.Source: Data availability · p. 28
  • Other included NICE guidelines covered chronic heart failure, adrenal insufficiency, cirrhosis, and Barrett’s oesophagus and stage 1 adenocarcinoma.Source: Data availability · p. 28
  • Additional NICE sources addressed pneumonia, asthma, falls, fertility problems, suspected sepsis, and hearing loss in adults.Source: Data availability · p. 28

Code availability

GUIDE’s source code is not publicly available at this stage because of safety concerns about unmonitored use in medical guideline workflows.

  • GUIDE’s source code is not publicly available at this stage because of safety implications for unmonitored medical guideline use.Source: Code availability · pp. 28–29
  • The editors and expert reviewers received the source code, testing dataset, and demonstration video as supplementary materials.Source: Code availability · p. 29
  • Project overview and deployment instructions are available through GitHub, while further details may be requested subject to institutional policies and approval.Source: Code availability · p. 29

using 20 datasets derived from four different specialties (cardiology, endocrinology,

This section describes GUIDE’s evaluation on post-cutoff and cross-specialty datasets alongside a prospective cataract evidence review, including its clinical questions, evidence extraction, and GRADE assessments of intraocular-lens comparisons. It also presents recommendations for individualized, informed shared decision-making in light of differing patient priorities and evidence limitations.

using 20 datasets derived from four different specialties (cardiology, endocrinology,

  • The 20-dataset, four-specialty evaluation assessed retrieval and screening, evidence-assessment consistency, recommendation agreement, and workflow time.The caption identifies cardiology, endocrinology, gastroenterology, and respiratory medicine; the supplied passages do not report quantitative outcomes for this evaluation.Source: using 20 datasets derived from four… · p. 40
  • 100.00% directional agreement with original NICE recommendations was achieved in the final step across six post-cutoff datasets, alongside consistently high screening recall and precision.The screening results were assessed despite potential upstream error propagation.Source: using 20 datasets derived from four… · p. 41, Figure 5 · p. 41
  • Two independent search agents generate retrieval strategies, and an AI Reviewer flags potential unreliability for expert-directed revision of Boolean strings.Experts give natural-language instructions through the assistant, which translates them into updated search strings.Source: using 20 datasets derived from four… · p. 42
  • Parallel screening agents display eligibility decisions, confidence, and rationales for expert review, with discrepancies prompting agent-specific reasoning for human adjudication.The workflow supports sequential title-and-abstract and full-text screening.Source: using 20 datasets derived from four… · p. 43
  • After full-text screening, GUIDE supports structured extraction of study-level text and outcome-level numerical results, with expert editing or re-extraction to verify discrepancies against the source.Extraction is organized around the PICO-defined evidence framework.Source: using 20 datasets derived from four… · p. 44
  • In the Delphi-like synthesis module, experts select a virtual panel whose members independently judge evidence and recommendations; a committee agent drafts a report and recommendation for expert revision and consensus.Panelists provide supporting rationales for their judgments and recommendation decisions.Source: using 20 datasets derived from four… · p. 45
  • The listed questions address corticosteroids for CAP or HAP and pulmonary-function or challenge-test monitoring of asthma; no performance results are provided here.Supplementary Table 8 is captioned as an overview of the simulated end-to-end prospective test dataset.Source: using 20 datasets derived from four… · pp. 58–59, Table 8 · p. 60

Supplementary Information 1. PICO criteria (prospective study, GUIDE-assisted group)

  • The GUIDE-assisted prospective review targets adults undergoing phacoemulsification and compares presbyopia-correcting intraocular lenses with standard monofocal lenses.Specified outcomes include distance and near visual acuity, spectacle independence, visual disturbances, additional procedures, satisfaction, quality of life, resource use, and costs.Source: Supplementary Information 1. PICO… · p. 61

GUIDE-assisted group)

  • The GUIDE-assisted evidence roster comprises prospective randomized comparisons of multifocal, trifocal, and extended-depth-of-focus lenses, including comparisons with monofocal lenses.Examples include trials of Vivity EDOF versus monofocal IOLs and PanOptix trifocal versus Vivity EDOF IOLs.Source: GUIDE-assisted group) · pp. 62–63
  • Six articles were excluded after human expert review, and cataract-only patient outcome data were marked not extractable.The supplied passages do not provide the full context for the separate note that an intervention did not meet criteria.Source: GUIDE-assisted group) · p. 68

Supplementary Information 4. GRADE tables (prospective study, GUIDE-assisted group)

  • The GRADE table lists 9 studies for corrected distance, uncorrected distance, and near-acuity thresholds, with 953, 943, and 1,043 participants, respectively.The corresponding effect estimates are truncated in the supplied passages.Source: Supplementary Information 4. GRADE… · p. 69
  • For mean corrected distance visual acuity, the GRADE table reports 9 studies and 1,265 participants; the effect estimate is truncated.The outcome is measured in logMAR, with lower values favoring multifocal lenses.Source: Supplementary Information 4. GRADE… · p. 69
  • Any spectacle dependence was lower with EDOF (RR 0.70, 0.59–0.85), but the evidence was rated very low certainty.The comparison included 8 studies and 985 participants.Source: Supplementary Information 4. GRADE… · p. 73
  • Halos were more frequent with EDOF (RR 2.08, 1.16–3.72), with low-certainty evidence.The comparison included 10 studies and 1,192 participants.Source: Supplementary Information 4. GRADE… · p. 74
  • Against monovision, multifocal lenses showed little difference in uncorrected intermediate visual acuity (MD 0.03, −0.04 to 0.10; moderate certainty) or near visual acuity (MD −0.03, −0.27 to 0.20; very low certainty).The intermediate acuity estimate came from 116 participants across 3 studies; the near acuity estimate came from the same 116 participants and 3 studies.Source: Supplementary Information 4. GRADE… · p. 76
  • In one 29-participant study, multifocal lenses showed no clear difference in maximum reading speed versus monovision (MD −5.00, −26.87 to 16.87; very low certainty).The outcome was measured in words per minute.Source: Supplementary Information 4. GRADE… · p. 76

GUIDE-assisted group)

  • No strategy was consistently superior across relevant outcomes, supporting individualized selection and informed shared decision-making rather than a routine preference for one option.The balance of benefits and harms is preference-sensitive; lack of a statistically significant difference does not establish equivalence.Source: GUIDE-assisted group) · p. 82
  • The committee advised assessing any lens price premium against benefits important to the individual and local opportunity costs, rather than assuming higher price means greater health benefit.Economic models were sensitive to assumptions about spectacle use, out-of-pocket costs, and willingness to pay; the cited studies were manufacturer-sponsored and model-based.Source: GUIDE-assisted group) · p. 83

Supplementary Information 4. GRADE tables (prospective study, GUIDE-assisted group)

  • Evidence certainty was predominantly low, with risk of bias, imprecision, small studies, and limited direct comparisons preventing a robust overall ranking.Evidence on patient-reported quality of life, dissatisfaction, serious complications, and additional treatment was less complete than evidence on visual acuity or spectacle use.Source: GUIDE-assisted group) · pp. 83–84

group)

  • Updated recommendations replace not offering multifocal lenses with considering them for people seeking greater spectacle independence.Source: group) · p. 85
  • Monovision is recommended for people who have used and want to continue with it after cataract surgery.Source: group) · p. 85
  • For multifocal lenses, recommendations call for discussing near and intermediate vision benefits and visual-disturbance risks, while considering ocular health, daily needs, and cost.Source: group) · p. 85
  • Listed comparison groups included refractive EDOF versus enhanced monofocal lenses and advanced monofocal versus EDOF lenses using mini-monovision.The passages also list comparisons involving multifocal and EDOF lenses and different refractive targets.Source: group) · pp. 88–89

Supplementary Information 10. GRADE tables (prospective study, human-only group)

  • RR 0.18 (0.08 to 0.41) for uncorrected near visual acuity worse than J3/J4 and RR 0.54 (0.40 to 0.74) for spectacle dependence favored multifocal lenses over monofocal lenses.Both estimates were rated low certainty.Source: Supplementary Information 10. GRADE… · pp. 92–93
  • RR 1.34 (1.06 to 1.83) for glare and RR 3.70 (1.89 to 7.23) for halos were reported for multifocal versus monofocal lenses.The table labels RR<1 as favoring multifocal; certainty was very low for glare and low for halos.Source: Supplementary Information 10. GRADE… · p. 94
  • MD −0.15 (−0.28, −0.02) for mean corrected intermediate visual acuity and RR 1.51 (1.03, 2.20) for halos were reported in the EDOF comparison.The table labels lower logMAR values and RR<1 as favoring EDOF; certainty was low for both estimates.Source: Supplementary Information 10. GRADE… · pp. 95–96
  • MD -0.03 (-0.08, 0.03) was reported for mean uncorrected distance visual acuity (logMAR) in the monovision-versus-monofocal table.The estimate was rated moderate certainty.Source: Supplementary Information 10. GRADE… · pp. 96, 98

human-only group)

  • Monofocal lenses provided greater distance acuity and fewer photic phenomena but often required near spectacles; greater monovision anisometropia could improve near vision while impairing binocular function.Prior successful contact-lens monovision may indicate tolerance, but caution was advised when precise depth perception matters.Source: human-only group) · pp. 103–104
  • Multifocal, especially trifocal, lenses generally offered the widest functional range, while EDOF lenses suited distance-to-intermediate tasks but did not reliably support small print or prolonged near work.Increased visual range could come with halos, glare, starbursts, or reduced contrast; EDOF lenses were not uniformly less disturbing than multifocal lenses.Source: human-only group) · p. 103
  • No IOL strategy was consistently preferable; recommendations favor individualized lens choice and explain that no strategy guarantees spectacle independence or symptom-free vision.Consider multifocal lenses when ocular health is suitable and patients highly value unaided intermediate and near vision with reduced spectacle dependence.Source: human-only group) · p. 104, Recommendations updated by human experts · p. 106
  • Affordability may constrain equitable access to advanced IOLs, but higher price should not be treated as evidence of greater clinical benefit.Local decisions should weigh patient-important benefits against incremental cost and update resource estimates as procurement prices and reimbursement change.Source: human-only group) · p. 105
  • Certainty was limited by serious risk of bias, inconsistency, imprecision, and small samples, reducing confidence that any strategy is superior overall.Evidence for monovision and several comparisons or patient-reported and adverse outcomes came from relatively few studies or a single trial.Source: human-only group) · p. 105

Consensus Interpretation

Loading 2610.16108v1…