Welcome back to Healthy Innovations! 👋

How many of us have pasted a blood test result into ChatGPT, Claude, or Gemini and asked what it means? If you haven't, you almost certainly know someone who has. Uploading personal health data into AI tools has become one of those behaviors that happened gradually – then all at once – driven less by hype and more by a simple gap: results arrive faster than explanations do.

This week's Deep Dive looks at what AI can and can't actually do with your blood test results, who the key players are, and why the distance between a chatbot and a clinically validated AI diagnostic is bigger than most people realize.

Let's dive in!

OpenAI has reported that tens of millions of people use ChatGPT for health-related questions every day. Recent KFF polling suggests roughly one in three US adults have used AI chatbots for health information in the past year, with a substantial share reporting they have uploaded personal health data, including blood test results, to get a personalized explanation.

Under the 21st Century Cures Act, most US health systems must release electronic health information without unnecessary delay, which means patients often see results before their clinician has reviewed them. The same shift is underway in the UK, where NHS patients can now view GP test results directly via the NHS App, often before their practice has been in touch. But even when a doctor does review first, a brief message saying "everything looks fine" doesn't explain why values are flagged or what they mean.

Either way, patients are left with questions. And increasingly, they're turning to AI to fill that gap.

Not all AI blood test interpretation is the same. There are two distinct use cases, and confusing them leads to confusing conclusions about what the technology can and cannot do.

Category 1: DIY interpretation. A patient receives results from their GP, an NHS portal, or a third-party testing service like Medichecks or Thriva, and pastes them into a general-purpose LLM (Claude, ChatGPT, Gemini) to get an explanation. It costs nothing beyond a subscription and requires no additional blood draw. In most cases, the AI has limited clinical context, no verified medical history, and no formal safety guardrails. It is working entirely from what the patient pastes in.

Category 2: Full-service platforms. Companies like Whoop and Levels handle the entire journey: blood draw or home kit, laboratory analysis, and AI-assisted interpretation of the results, typically with a clinician review layer on top. These are subscription products, some costing hundreds of dollars per year, that pitch themselves as a smarter, more personalized alternative to a standard annual physical.

The clinical evidence gap applies to both. But it looks different in each case.

Early, informal evaluations of general-purpose AI models on real blood test scenarios suggest they frequently miss important follow-up steps and inconsistently advise patients to seek clinical care. There is currently no large, standardized benchmark for this kind of AI blood test interpretation, which makes it almost impossible to compare tools or verify claims.

In March 2026, Quest Diagnostics launched its AI Companion, powered by Google's Gemini models, specifically to help patients understand lab terminology and identify trends in up to five years of their Quest results. Quest drew a deliberate line: the tool explains your results. It does not give medical guidance or lifestyle advice.

That restraint is notable. Most general-purpose LLMs have no such guardrails when a patient pastes in a lab report. The quality of the response depends entirely on the model, the prompt, and how much context the patient provides.

Whoop and Levels represent the more structured end of the consumer market. Both combine blood testing with wearable data, AI analysis, and clinician-reviewed reports. Both use LLMs trained on medical literature, with Whoop also integrating physiological data such as activity and sleep to add context to results.

Neither company has published peer-reviewed research demonstrating the accuracy of its AI interpretation product. Founders in this space openly acknowledge there are no widely accepted benchmarks for comprehensively interpreting blood tests at scale. Some are planning validation studies through health system partnerships, but most of those efforts are still forthcoming.

Several full-service platforms also market "clinician-reviewed" reports as a quality signal. But researchers, including those at Mount Sinai who have published on AI in clinical decision-making, have raised concerns about automation bias: in some studies, clinicians over-trust AI outputs and approve them without independent scrutiny. Physician oversight is necessary but not sufficient.

Both categories sit a long way from what validated AI blood testing looks like at the clinical end of the spectrum.

In late 2025, Nature Medicine published results from the SEPSIS-SHIELD trial, a prospective study of TriVerity, an AI-based blood test developed by Inflammatix. Tested across multiple emergency departments in the US and Europe, enrolling 1,222 patients, TriVerity analyzes 29 immune-related mRNAs from a blood sample to determine the likelihood of bacterial infection, viral infection, and the need for critical care within seven days.

The results were notable:

As a result, in January 2025, TriVerity received FDA 510(k) clearance – the standard US regulatory pathway for medical devices, roughly equivalent to a CE mark in the UK. Clearance requires the manufacturer to demonstrate the device is safe and effective before it can be sold.

This is what validated AI blood testing looks like: a prospective trial, published in a peer-reviewed journal, tested across diverse populations, with regulatory clearance earned through a formal submission.

The research draws the same broad lines across both categories.

Where AI adds real value:

Where the evidence runs out:

Experts and professional bodies have repeatedly made this point: there is no rigorous published research demonstrating that AI can accurately make personalized health recommendations from blood results alone. That may change, but it has not changed yet.

A further concern applies to both categories.

A 2025 Nature Medicine study from the Icahn School of Medicine at Mount Sinai stress-tested nine general-purpose LLMs on 1,000 emergency department cases and found that recommendations shifted based on patients' race, housing status, income, and sexual orientation, even when the clinical picture was identical.

A follow-up study by the same group, published in Nature Medicine in May 2026, found that ChatGPT Health – OpenAI's purpose-built consumer health tool – showed more resistance to demographic bias, suggesting that purpose-built health tools can partially mitigate these effects. But new safety concerns emerged: ChatGPT Health undertriaged 52% of true emergencies, directing patients with diabetic ketoacidosis or respiratory failure to routine care rather than the emergency department.

The lesson from both studies is the same: consumer-facing AI health tools, however they are designed, need prospective safety validation before widespread deployment.

Any conversation about blood testing has to acknowledge the shadow of Theranos. Elizabeth Holmes built a company on the promise of revolutionary diagnostics from a few drops of blood, raised enormous investor capital, and went to prison when it emerged the technology never worked as claimed. The failure was not just fraud. It was the habit of moving fast in a domain that demands the opposite.

Today, Haemanthus, a new startup led by Holmes' partner Billy Evans, is raising capital for an AI-powered diagnostic device using photonics to analyze bodily fluids. According to PharmaVoice, the company has acknowledged the comparison directly, stating publicly that skepticism is rational and that it must clear a higher bar. No peer-reviewed data or clinical validation has been published to date.

The lesson is that extraordinary claims require peer-reviewed evidence – and the startups that will matter are those doing the validation work now.

Right now, the most honest framing for AI blood test interpretation – in either category: it is a knowledgeable first read, not a diagnosis.

The technology will improve. Benchmarks will emerge. Some of these platforms will do the hard work of clinical validation and earn the trust they are currently asking for. But right now, AI blood test interpretation is most valuable as a tool for understanding and advocacy, helping patients arrive at clinical conversations better prepared, with better questions.

That is a genuine contribution to healthcare. It is also a much smaller claim than most of the marketing makes.

Understanding how your body uses fuel – whether it's trending toward fat or carbohydrate use at any given time – has traditionally required a lab. Lumen brings that closer to your kitchen counter.

The handheld breath device measures CO₂ in your exhale to estimate whether your body is trending toward fat or carbohydrate use. The connected app pairs those readings with nutrition and lifestyle guidance, adjusting recommendations based on your activity, sleep, and recent results. Over time, it tracks your metabolic flexibility: how efficiently your body switches between fuel sources, which matters for energy, weight management, and long-term health.

Founded by twin sisters with PhDs in physiology, peer-reviewed studies suggest Lumen can detect meaningful changes in metabolic fuel use and post-meal responses – though it remains an indirect consumer device rather than a medical diagnostic tool. Useful for anyone who wants to move beyond generic nutrition advice and start tracking how their metabolism actually responds.

🫏 Donkeys on the hospital payroll. At a psychiatric hospital outside Paris, five therapy donkeys – Nono, Pitou, Oscar, Manolo, and Malraux – are doing something no antipsychotic has quite managed: getting patients to leave their rooms with a smile.

Running since 2016 at the Ville-Evrard hospital complex in Neuilly-sur-Marne, the program pairs patients living with anxiety, depression, autism, and schizophrenia with donkeys described by their trainer as "emotional sponges." Staff report meaningful improvements in confidence, communication, and self-worth. The program has since expanded to guinea pigs, chickens, goats, turtles, and rabbits – essentially a farmyard with clinical governance.

It now holds official healthcare unit status with three full-time nurses. The only thing still missing? Peer-reviewed evidence. Though one nurse may have settled the debate: "Donkeys are my best colleagues."

Thank you for reading the Healthy Innovations newsletter!

Keep an eye out for next week’s issue, where I will highlight the healthcare innovations you need to know about.

Have a great week!

Alison ✨

P.S. If you enjoyed reading the Healthy Innovations newsletter, please subscribe so I know the content is valuable to you!