📄 nutrition

Cross-LLM AI platform meta-research: Non-inferiority of bovine milk-based fortifiers to human milk-based fortifiers

This study presents a proof-of-concept cross-LLM AI platform meta-research that, by analyzing 3,371 publications using ChatGPT, Claude, and Manus AI, concludes that bovine milk-based fortifiers are non-inferior to human milk-based fortifiers for preventing necrotizing enterocolitis and sepsis in pre-term newborns, thereby demonstrating the viability of AI-assisted evidence synthesis in medicine.

Original authors: Ni, D., Ge, A., Mishra, A., Oei, J. L., Nanan, R.

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ni, D., Ge, A., Mishra, A., Oei, J. L., Nanan, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: A High-Stakes Baby Food Showdown

Imagine a premature baby is like a very delicate, high-performance race car that needs special fuel to grow strong. Doctors have two main types of "fuel additives" (fortifiers) to mix into the baby's milk:

Human Milk-Based Fortifiers (HMF): Made from human milk. These are the "premium, organic" option. They are expensive and many people believe they are the absolute best at protecting the baby's gut and preventing serious infections.
Bovine Milk-Based Fortifiers (BMF): Made from cow's milk. These are the "standard, reliable" option. They are much cheaper.

For years, there has been a heated debate in the neonatal intensive care unit (NICU): Is the expensive premium fuel actually better than the standard fuel? Specifically, does it stop a dangerous gut condition called Necrotizing Enterocolitis (NEC) and blood infections (sepsis)?

The problem is that previous studies have been messy. They are like trying to compare two cars by driving them on different tracks, with different drivers, and different weather. The results have been confusing, and some studies might have been influenced by the companies selling the expensive fuel.

The New Approach: The "AI Panel of Judges"

Instead of hiring a team of human researchers to spend years reading thousands of papers (which is slow, expensive, and prone to human bias), the authors of this paper tried something new. They built a "Cross-LLM AI Platform."

Think of this as hiring three independent, super-smart AI judges (ChatGPT, Claude, and Manus AI) to act as a panel.

The Task: They fed these AI judges a massive library of 3,371 scientific papers about baby milk fortifiers.
The Rules: The AI judges were given a strict, standardized checklist (like a referee's rulebook) to find only the fairest, most direct comparisons (Randomized Controlled Trials) where the only difference between two groups of babies was the type of fortifier they got.
The Goal: To see if the AI judges could agree on the answer without getting tired, distracted, or influenced by who paid for the studies.

What the AI Judges Found

After the three AI judges read through the best available evidence, they all raised their hands and agreed on the same verdict:

The "Premium" fuel (Human Milk) is not statistically better than the "Standard" fuel (Cow's Milk) for preventing NEC or sepsis.

The Gut (NEC): The AI found that babies fed the cow-milk-based fortifier had the same rate of gut problems as those fed the human-milk-based one. The expensive option didn't offer extra protection.
The Infection (Sepsis): Similarly, there was no difference in infection rates between the two groups.

It's as if the AI judges looked at the race cars and said, "Even though the premium fuel costs three times as much, both cars crossed the finish line with the same number of breakdowns."

Why This Matters (According to the Paper)

Solving the Debate: The paper claims this provides strong, unbiased evidence that the expensive human-milk fortifiers are non-inferior (just as good as) the cheaper cow-milk ones for these specific outcomes. This could save hospitals and families a lot of money.
A New Way to Do Research: The authors aren't just talking about baby food; they are proving a concept. They showed that using multiple AI tools to cross-check each other is a fast, transparent, and reliable way to do "meta-research" (research about research).
- Analogy: Usually, doing a big review of all medical studies is like trying to count every grain of sand on a beach by hand. This paper suggests using three different high-tech sand-sifting robots that agree with each other is a faster, more honest way to get the count.

Important Caveats (What the Paper Does Not Say)

It's a Preprint: The paper explicitly states this is a "preprint," meaning it hasn't been officially peer-reviewed by other scientists yet. It is a proof-of-concept, not a final medical rule.
No New Clinical Advice: The paper does not tell doctors to stop using human milk fortifiers. It simply says the evidence doesn't prove they are better than cow milk for these specific issues.
AI Limitations: The authors acknowledge that AI can make mistakes (hallucinate), but they argue that using three different AIs and having humans check the work minimizes this risk.

In summary: This paper uses three AI "detectives" to solve a decades-old mystery in baby care. Their conclusion? The expensive human-milk additives aren't doing a better job than the cheaper cow-milk ones at preventing gut damage and infections. The paper also argues that using AI to review medical evidence is a promising, fast, and fair new way to do science.

Technical Summary: Cross-LLM AI Platform Meta-Research on Bovine vs. Human Milk-Based Fortifiers

Problem Statement
Necrotizing enterocolitis (NEC), often leading to sepsis, remains a primary cause of morbidity and mortality in preterm neonates. While human milk is the standard preventive approach, premature infants often require fortification to achieve adequate growth. A significant clinical and economic dilemma exists regarding the choice between human milk-based fortifiers (HMF) and bovine milk-based fortifiers (BMF). Although HMF is widely presumed superior, robust evidence supporting this claim is lacking due to heterogeneous study designs, mixed dietary interventions, and a scarcity of direct comparisons. Furthermore, HMF imposes a substantial financial burden on healthcare systems. Traditional systematic reviews and meta-analyses, the gold standard for evidence synthesis, are criticized for being time-consuming, resource-intensive, and susceptible to human bias and subjectivity, particularly in fields with potential conflicts of interest.

Methodology
The authors propose and execute a "cross-LLM AI platform meta-research" workflow to synthesize evidence and cross-validate findings without relying solely on human-driven processes.

Data Curation: A literature search on PubMed (conducted May 7, 2026) yielded 3,371 publications related to human milk fortifiers. Among these, 130 were identified as systematic reviews, meta-analyses, or network meta-analyses.
AI Platforms: Three state-of-the-art Large Language Model (LLM) platforms were utilized in parallel: ChatGPT (GPT-5.5), Claude (Opus 4.7), and Manus AI (Manus Pro, Manus 1.6).
Workflow:
1. Retrospective Analysis: The AI platforms were prompted to screen the 130 existing systematic reviews to identify those directly comparing HMF and BMF regarding NEC and sepsis.
2. Primary Evidence Synthesis: The platforms were tasked with identifying Randomized Controlled Trials (RCTs) from the 3,371 publications that directly compared HMF vs. BMF.
3. Data Extraction & Analysis: For the selected RCTs (11 papers with full text available), the AI platforms were provided with the full-text manuscripts, supplementary data, and the "Revised Cochrane risk-of-bias tool for randomized trials (RoB 2)." Using standardized prompts, the AI extracted data and performed random-effects meta-analyses independently.
4. Cross-Validation: Results from the three AI platforms were compared against each other and reviewed by human authors to ensure consistency and mitigate the risk of AI hallucinations.

Key Contributions

Proof-of-Concept Workflow: The study introduces a novel, reproducible methodology for evidence synthesis using multiple LLM platforms to cross-validate findings, aiming to reduce human bias and increase transparency.
Direct Head-to-Head Comparison: By isolating RCTs where the fortifier type was the sole variable, the study provides a cleaner comparison between HMF and BMF than previous mixed-intervention reviews.
Conflict of Interest Analysis: The workflow successfully identified and highlighted conflicts of interest (including industry funding) in both prior systematic reviews and the primary RCTs, suggesting potential sources of bias in existing literature.

Results

Existing Literature Review: The AI analysis of 10 identified systematic reviews revealed mixed findings: 4 reported HMF superiority for NEC, while 6 found no difference. For sepsis, most reviews indicated non-inferiority of BMF. Two of these reviews disclosed financial conflicts of interest with HMF manufacturers.
Meta-Analysis of RCTs:
- NEC Outcomes: All three AI platforms consistently found no statistically significant difference between HMF and BMF.
  - Manus AI (3 studies): RR 0.78 (95% CI 0.37–1.63), $I^2 = 0.0\%$ .
  - ChatGPT (4 studies): RR 0.78 (95% CI 0.37–1.63), $I^2 = 0.0\%$ .
  - Claude (5 studies): RR 0.75 (95% CI 0.37–1.51), $I^2 = 0.0\%$ .
- Sepsis Outcomes: Similarly, no evidence was found that HMF is superior to BMF.
  - Manus AI (4 studies): RR 1.11 (95% CI 0.57–2.14), $I^2 = 45.1\%$ .
  - ChatGPT (5 studies): RR 1.04 (95% CI 0.58–1.85), $I^2 = 32.1\%$ .
  - Claude (5 studies): RR 1.04 (95% CI 0.62–1.77), $I^2 = 38.0\%$ .
Conclusion: The synthesis demonstrates the non-inferiority of BMF to HMF for both NEC and sepsis outcomes.

Significance and Claims
The paper claims to provide independent evidence addressing a critical controversy in neonatology, suggesting that the more expensive HMF does not offer superior protection against NEC or sepsis compared to BMF. Beyond the clinical findings, the authors position this work as a methodological breakthrough. They argue that cross-LLM AI workflows offer a faster, more transparent, and less biased alternative to traditional systematic reviews. The study asserts that this approach can be extended to other research questions and AI models (e.g., Gemini, DeepSeek), representing a new paradigm for integrating AI into evidence-based medicine and digital health. The authors explicitly state that this is a proof-of-concept and that their findings should be interpreted within the context of the specific RCTs analyzed, without claiming the superiority of the AI platforms themselves over human reviewers, but rather their utility as a complementary, unbiased tool.