If you follow health influencers or longevity experts online, you’ve likely heard about Nattokinase, an enzyme derived from fermented soybeans.

It’s frequently hyped as an inexpensive miracle supplement that can single-handedly melt away up to 95% of arterial plaque, offering a cheap alternative to expensive statins or invasive surgical procedures.

Influencers routinely point to scientific literature to back up these bold claims. But does the published research actually say what they claim it does?

To separate online hype from scientific reality, we ran a Tessa analysis on one of the primary systematic reviews frequently cited to validate nattokinase: “Nattokinase Supplementation and Cardiovascular Risk Factors: A Systematic Review and Meta-Analysis of Randomized Controlled Trials.“

Here is what Tessa revealed when we put this paper under the microscope.

What the Paper Actually Showed

Despite social media claims that nattokinase clears plaque and dramatically lowers cholesterol, this meta-analysis of randomized controlled trials (RCTs) found something very different:

• No Plaque Busting or Lipid Lowering: Across the pooled trial data, nattokinase showed no consistent lipid-lowering effect. In fact, lower total doses of nattokinase were actually associated with a slight increase in total cholesterol and LDL (“bad”) cholesterol, and a slight decrease in HDL (“good”) cholesterol.
• Modest Blood Pressure Reduction: The paper did confirm a small, statistically significant drop in blood pressure (roughly 3.4 mmHg systolic and 2.3 mmHg diastolic). While helpful, this modest drop is a far cry from replacing standard medical treatments.
• A Slight Increase in Blood Glucose: The analysis noted a small but statistically significant increase in fasting blood glucose among participants taking nattokinase.

The Tessa Overall Score

TScore: 66 / 100 ( Yellow – Caution)

A graphical illustration of Tessa’s analysis of the paper

Tessa Analysis Results

Tessa Sub-Scores

The Tessa Analysis Breakdown

While the authors followed standard PRISMA reporting protocols (achieving an 88% completeness score), Tessa Pro’s deep-dive evaluation highlighted major structural vulnerabilities in the underlying evidence:

• Methodological Quality (AMSTAR 2): Critically Low Confidence — Although the authors conducted duplicate screening and used standard risk-of-bias tools, the review suffered from critical methodological gaps. Key issues included an incomplete search strategy (excluding trial registries and grey literature), failure to provide a detailed list of excluded studies with justifications, and no reporting on the funding sources of the primary trials analyzed.
• Certainty of Evidence (GRADE): Very Low Certainty — When evaluated using the GRADE framework across 8 primary outcomes, the certainty of evidence dropped to Very Low for most findings, including LDL, HDL, blood pressure, and adverse event reporting. The primary driver for this rating was the small number of overall participants (only 6 RCTs totaling 546 subjects) combined with unexplained heterogeneity (variability) between trial results.
• Citation Integrity: Tessa’s citation engine verified 95% of the paper’s references with an average relevance score of 7.1/10, showing solid context alignment without evidence of aggressive citation padding.

The Bottom Line: While the systematic review itself was conducted in good faith, the actual primary data simply does not support the massive health claims circulating on social media.

See for yourself: Read the full, objective Tessa analysis report here (includes links to the study): https://www.tessapp.ai/report/39076715

If you’re interested in similar studies, check out the Tessa analysis of these:

• Cohort Study (1,062 people): (65 / 100 ( Yellow) — Effective management of atherosclerosis progress and hyperlipidemia with nattokinase: A clinical study with 1,062 participants. (Frontiers in cardiovascular medicine)
• Narrative Review: (50 / 100 ( Yellow) — Research Progress of Nattokinase in Reducing Blood Lipid. (Nutrients)

Our first full week of the Tessa public preview has revealed a fascinating trend: the public is deeply concerned about how plastic waste impacts our world, with a massive surge of interest in research surrounding microplastics in our food supply.

To tackle the plastic crisis, governments worldwide are pushing hard to ramp up recycling and composting. But walk into any kitchen or cafeteria, and you’ll find the real bottleneck: human confusion. Most consumers can’t reliably tell the difference between non-recyclable, recyclable, and compostable plastics. When everything gets tossed into the same bin, recycling facilities are forced to resort to slow, incredibly expensive manual sorting.

The unfortunate result? Tons of perfectly salvageable material gets dumped straight into landfills anyway.

A study published last summer by researchers at University College London (UCL) aims to fix this exact issue. Titled “Using hyperspectral imaging and machine learning to identify food-contaminated compostable and recyclable plastics,” it explores an AI-driven method to sort the mess for us.

We ran this paper through Tessa to see how well its promises hold up.

The Tessa Verdict: A Verified Green Light

Overall T-Score: 72/100 (Green )

Tessa Insights: A score of 72 officially crosses the threshold into our Green zone. While it sits at the lower end of the green spectrum, remember that perfect science does not exist. A 72 actually places this UCL study in the top 20% for quality among papers evaluated, making it a highly credible and reliable piece of literature.

The Deep Dive Breakdown

• From Theory to Reality (Score: 90/100): This paper is highly experimental and hands-on. Instead of relying purely on clean computer models, the UCL team got their hands dirty by applying actual food contaminants (like ketchup and mayonnaise) to physical plastic samples. They even validated their system on a small batch of 30 real-world waste items. It only misses a perfect score because it hasn’t been tested in a live, high-speed commercial facility yet.
• Practical Innovation (Score: 60/100): Using advanced cameras and AI to identify materials is an established concept. However, applying it to the specific, messy problem of food-covered packaging adds a highly practical, timely upgrade to the field. It builds smartly on existing tech to solve a real-world headache.
• Scientific Rigor & Speed Bumps (Score: 50/100): This is the main reason the paper sits at a low green rather than a high one. While the researchers reported a success rate of over 90% (up to 99% in specific settings), their real-world test size was quite small (only 30 items). Additionally, the data analysis leaves room for overly optimistic accuracy estimates depending on whether the AI looked at individual pixels or the whole object. It also lacked standard statistical safety checks to prove absolute consistency.

Why It Matters

The UCL team has proven that advanced imaging can see past the ketchup stains to accurately identify what kind of plastic is underneath. Because it has earned a green score from Tessa, the foundational science is solid. While we still need to see it tested on a much larger scale before it hits local waste plants, this research provides a vital, high-quality blueprint for automated sorting.

Want to inspect the data yourself? Explore the full Tessa Report here.

Artificial Intelligence has tremendous potential, but there’s a long way to go between what might be possible one day and what is real today.

AI isn’t quite up to the task yet

As the founder of a tech startup, I spend a lot of time talking about my company and the problem it’s trying to solve. Naturally, because artificial intelligence is such a hot topic these days, people assume that it’s a big part of the “magic” in our platform’s technology.

The truth is though, AI only plays a minor role. It’s not that it isn’t a useful technology, it’s just that, despite what the hype is telling you, it’s not a good solution for every type of problem. Actually, it’s not a good solution for most problems, particularly the complex ones in medical and scientific research we’re trying to solve.

Even for those problems that AI is good at, it still has a long way to go before the real value of AI matches the hype. Here are the key weak spots we see in AI today:

AI promises are overblown

Despite the high-profile entrepreneurs trying to create a wave of excitement and raise money to pursue their ventures, most of the promises and predictions are really based on hopeful thinking rather than reality.

Yes, it would be nice if AIs were like Tony Stark/Iron Man’s J.A.R.V.I.S. – conversational, fast, and 100% accurate. But the reality is, today’s AIs, even after extensive training, might get to 90% accuracy… most of the time. With additional, highly focused training, it might get to 95%. But that last 5% will probably never happen.

Admittedly, 95% is pretty good if you’re working on a 5th grader’s book report, or trying to learn how many bluefin tuna are caught off the coast of Mexico each year. But when it comes to subjects where “good enough” isn’t – and there are all kinds, ranging from aircraft design to life-saving biopharma research – it’s nothing you’d want to trust.

AI is power hungry and expensive

Finally, AI is incredibly expensive. Running an AI requires immense computational resources which, in turn, requires immense amounts of electricity to power, heat and cool the hardware. Even big tech companies like OpenAI, Microsoft, and Meta publicly admit those costs are so high that their AI initiatives struggle to be profitable. Additionally, the more complex the AI, the more expensive it is to run.

Typically, we would expect the costs of technology to decline over time. However, the real costs of AI aren’t in the software or hardware, they are in the power and bandwidth it takes to run it. Demand for AI is so high that global data center capacity, which will double over the next five years, will still lag demand. Moreover, electricity consumption is expected to increase almost 25% a year through the end of the decade. That double-whammy is expected to keep prices sky high through at least the end of the decade.

AI isn’t good at most tasks

Beyond the hype problem and expense, there are simply many kinds of real-world problems that AI just can’t solve.

A lot of the appeal in AI is the idea that it’s some sort of super smart Star Trek computer that can answer questions and solve problems better, faster and cheaper than a human. However, AI isn’t actually “intelligent”, it’s just a really big statistical analysis system. It’s very good at giving you the most likely answer based on a statistical match with the questions you ask it, but it’s completely useless when it comes to “understanding” the question.

For example, if you show a person a “no trespassing” hanging upside down and ask them “can you fix that?” the person will have no problem recognizing that the sign is upside down and turn it right side up. An AI on the other hand, unless it has been specifically trained on that upside down imagery, and that for it to be proper it must be turned 180 degrees, will fail every time.

In other words, AIs can’t reason or do novel problem solving the way people can, so their use is limited to repeating that which has already been done. Again, this is fine for the 5th grade book report on Moby Dick, but when it comes to finding novel solutions or answering questions that haven’t already been answered and documented, AI is neither better nor cheaper than a human being.

Where does AI fit?

Having listed a number of ways AI isn’t the end-all-be-all technology, I should note where AI does have its uses.

As I mentioned above, AI can’t create anything that hasn’t been created before, but it is very good at (re)processing things that already exist to present them in a new way. For example, AI is very good at re-stating and summarizing longform written content. It’s also pretty good at identifying similarities and relationships among different content different pieces. And if you want to know what the pope would look like in a puffy Balenciaga jacket, AI is the way to go.

At Siensmetrica, we’re using AI to quickly deliver summaries of research papers, identify the points of consensus among distributed findings, and even help filter results among well-documented metrics like the h-index.

But when it comes down to diving deep into materials to identify novel or difficult-to-articulate measures of what makes the content of one thing more likely to be reliable than another, we’re relying on a whole series of inputs, multi-vector data analysis, and fuzzy weighting factors like trust, explainability and significance to do it.

Maybe one day AI will have the capabilities and the cost-benefit ratio that make it a viable all-in-one solution for the technology we’re developing. But until then, we’re not going all in on it.