Systematic Reviews Hub Β· Cochrane-Level Β· Built in Public

Clinical AI, pooled and graded.

Three systematic reviews and meta-analyses by Raphael Malikian β€” healthcare founder and operator, now building clinical AI. Each review was carried out by a multi-agent LLM team filling the seats a Cochrane team would fill, with one accountable human author signing off every number. Protocol, search log, screening calls, extraction sheets, analysis code and GRADE ratings are public for every review. Check the work, do not just trust it.

3
Reviews complete
148
Studies included in total
18,500+
Records screened
3
Zenodo DOIs Β· Free, CC BY 4.0
$0
Funding Β· No conflicts

These are Cochrane-level reviews. They are not Cochrane reviews and were not submitted to Cochrane. Certainty is honestly graded β€” and for most outcomes in this young literature, it is low or very low. That honesty is the point.

THE REVIEWS

Read them in order, or start with the question your organisation is spending money on.

REVIEW 1 Β· 22 OCT 2026AMBIENT AI SCRIBESSUBMITTED β€” J CLINICAL EPIDEMIOLOGY

Ambient AI scribes for clinical documentation: documentation time, after-hours work, clinician burnout, note quality and note errors

1,049 records found Β· 567 screened after deduplication Β· 74 studies included

Scribes cut documentation time by about a fifth β€” and 59.8% of audited AI-drafted notes contained at least one error. The time saving and the checking burden are the same intervention.
βˆ’18.9% documentation timePooled percentage change, 15 studies Β· GRADE low
SMD βˆ’0.44 work exhaustion2 randomised trials, 30–90 days Β· GRADE moderate for the trial estimate
18.0% omissions Β· 11.5% hallucinationsOne audit of 356 notes against the consultation record Β· GRADE very low
ΞΊ 0.64 screening agreementDual independent screening, disagreements adjudicated in the open
REVIEW 2 Β· 22 OCT 2026 Β· ROBUSTNESS PASS 23 OCTLLM CITATION FABRICATIONRESEARCH INTEGRITY

Fabricated and inaccurate citations generated by large language models for biomedical topics

16,474 records screened Β· 63 generation experiments Β· 148 model arms Β· 22,461 references checked one by one

About 1 in 5 LLM-generated biomedical references did not exist at all β€” 19.4% completely fabricated, only 37.0% fully accurate. Turning retrieval on cut fabrication to 5.8%, versus 31.0% with retrieval off.
19.4% fabricated95% CI 14.9–24.8 Β· 113 arms, 50 studies Β· GRADE very low
37.0% fully accurate95% CI 27.2–47.9 Β· GRADE very low
16.5% any major inaccuracy95% CI 5.7–39.2 Β· GRADE very low
5.8% vs 31.0%Fabrication with retrieval on vs off β€” verify every reference
REVIEW 3 Β· 31 OCT 2026LLM CLINICAL SUMMARIESREPORTING GAP FINDING

Clinically significant errors in LLM-generated clinical summaries: omissions, inaccuracies and fabrications per summary

991 records found Β· 483 screened after deduplication Β· 11 studies included Β· 22 arms Β· All summary types, discharge as subgroup

About 3 in 5 LLM summaries contained at least one error of any audited type β€” yet only 1 of 11 studies reported clinically significant errors per summary. In that one study, 21.0% of summaries carried a critical, treatment-impacting error. The literature rarely counts the error that matters most.
60.6% any errorSupplementary pool, 5 units, 323 summaries Β· GRADE very low
20.7% omissions Β· 27.9% inaccuraciesPooled per-summary proportions Β· GRADE very low
17.6% fabrication / additionRange 0–97.5% across studies β€” the spread is the finding Β· GRADE very low
Primary outcome: not pooledOnly 1 study reported it per summary β€” a formal reporting-gap result
NEW Β· 21 NOV 2026VIRAL TASTE LOSSFAMILY MEDICINE

Loss of taste in possible viral infection: etiologies, pathogenesis, and the evidence for treatments

5,988 records found Β· 3,501 screened after deduplication Β· 363 studies included Β· 362 unique publications

About 38 in 100 people with COVID-19 in these studies had taste dysfunction early in the illness β€” and every pooled study concerns SARS-CoV-2, not flu or any other named virus. Most people's taste came back fully. Only 23 studies tested any treatment, and none can be recommended for or against.
37.7% acute taste dysfunctionPooled, 137 studies, 155,010 participants Β· 95% CI 32.9–42.7 Β· GRADE very low
85–88% complete recoveryAt every window studied, from different study sets Β· GRADE very low
73.4% of mechanism evidence speculative214 studies with a mechanism code Β· only 15 with direct human evidence
23 treatment studies Β· 2 RCT taste estimatesPrednisolone null in the best trial Β· no recommendation made

HOW THEY WERE BUILT

The method is part of the claim β€” so the method is now open source too. A multi-agent team, strict independence rules, and a shipping gate that failed builds do not pass.

πŸ‘₯ A team, not a prompt

Information specialist, dual screeners who never see each other’s calls, adjudicators who settle disagreements by re-reading, dual extractors, a risk-of-bias pair, statistician, GRADE assessor, methods editor and a separate visual-QA pass.

πŸ” Nothing invented, nothing hidden

No invented numbers, counts or citations. A source that was not run is never reported as searched. Blocked sources, discrepancies and corrections stay visible in the team log.

πŸ“Š Checked two ways

Random-effects models named in advance, heterogeneity and prediction intervals reported, at least one pool hand-checked in full, and code + data published so anyone can rebuild every figure.

πŸ–ΌοΈ Figures inspected, not assumed

Every forest plot carries a summary diamond. Every figure and table page is rendered and visually inspected before shipping. A build that fails inspection does not ship.

Use the open process β€” prompts, templates & Robust Standard v2 β†’   Copy the example multi-step prompt β†’

Process licence: free for non-profit use with acknowledgement of Raphael Malikian. Commercial use requires a licence β€” email rtmalikian@gmail.com. The three reviews themselves are CC BY 4.0.

CITE & CONTACT

Raphael Malikian Β· Independent researcher, Palmdale, CA, USA Β· ORCID 0009-0008-4326-4205 Β· rtmalikian@gmail.com Β· GitHub @rtmalikian Β· X @rtmalikian

Cite each review by its Zenodo DOI above. Funding: none, for all three reviews. Conflicts of interest: none. No industry affiliation, no vendor support. Scope: evidence synthesis and policy discussion only β€” no clinical or medication recommendations.