AI Can Predict Disease Years Early
with Mariano, Axenya
AI Can Predict Disease Years Early
Show Notes
Mariano is a four-time healthcare entrepreneur and CEO of Axenya, an AI-powered operating system for corporate health that monitors over 100,000 people and runs 95 million clinical inferences per month. In this conversation, he breaks down why a decade of promises about transforming healthcare with technology haven't delivered - and what is finally different now. The answer involves two things arriving at the same time: the ability to collect continuous health data at scale, and AI capable of making sense of it.
This episode goes deep on the systemic incentive problems that keep healthcare broken, why Axenya designed its entire business model around outcome-aligned payment, how AI is functioning as a copilot for clinicians rather than a replacement, and what the future of preventive medicine looks like when you can flag someone as high risk years before they develop symptoms.
The Chronic Disease Problem - and Why Traditional Healthcare Can't Solve It
Modern healthcare was designed in the 1940s to treat acute diseases - infections, injuries, conditions that present with clear symptoms and respond to targeted interventions. That model worked. Then the epidemiological landscape shifted to chronic disease, and the model didn't.
Chronic diseases - diabetes, hypertension, cardiovascular disease - share two properties that break the traditional model. First, they have almost no symptoms until very late. Symptoms are the signal that drives people to seek care; when the signal is absent, people don't go. Second, they require long, complex, feedback-poor management over years. The same visit-based, reactive system that works for acute illness is structurally unsuited to catching and managing diseases that are quietly progressing without announcing themselves.
Mariano's thesis, developed five years ago before the current AI wave: two new capabilities could change this entirely. Continuous data collection - measuring blood pressure not just at the doctor's office but throughout the day - and AI capable of analyzing dozens of data points per patient per minute across a whole population. Together they create the possibility of finding patients before they're sick enough to seek care on their own.
Layered Data and 95 Million Inferences a Month
Axenya ingests three primary categories of data. Financial data - from the insurance system - is the cleanest: every pharmacy purchase, every doctor visit, every lab order leaves a payment trail. Clinical data comes from wearables (Apple Watch, glucose monitors, continuous heart rate), from a 30-second face scan video that uses an MIT-developed model to infer 32 variables including blood pressure and cholesterol from facial blood vessel coloration, from questionnaires, from lab results patients share directly, and from mandatory annual occupational health exams. Partner data flows in from third-party programs - obesity management, mental health platforms - that integrate with the system.
No single source is sufficient. Wearable penetration in Brazil is around 15% of the population. Face scans provide inference, not diagnosis. Labs are episodic. The power is in combination: when financial, clinical, and behavioral data are fused into a single data lake and analyzed continuously, patterns emerge that no individual data source could produce alone. The result is 95 million clinical inferences per month for 100,000 patients - a coverage density that would require the entire physician population of Brazil to replicate manually.
The Incentive Problem - and the Business Model That Solves It
The structural failure of healthcare isn't primarily technological. It's incentive misalignment at every layer. Hospitals make more when patients receive more treatment. Insurers who invest in prevention often don't hold the policy long enough to collect the financial benefit - the average corporate insurance contract lasts one year, and chronic disease prevention plays out over five to ten. So the rational behavior for everyone in the current system is to treat, not prevent.
Axenya resolved this by redesigning their revenue model from the ground up. They don't charge for the digital tools. They don't charge a per-seat SaaS fee. They charge a performance fee based on the delta between a client's healthcare cost growth and the market baseline - which in Brazil runs at 2.5x inflation. If Axenya lowers cost growth, they earn. If they don't, they don't. 95% of their clients have been saving money for three consecutive years. The technology is sophisticated; the business model is what makes the technology fundable and deployable.
Mariano's broader point is that this alignment is a prerequisite for trustworthy AI in regulated industries, not just a clever pricing strategy. Any system that profits from the disease it claims to treat - or that isn't accountable to the outcomes it promises - will optimize for the wrong thing. The revenue model is the accountability mechanism.
AI as Copilot - Transparent Reasoning Required
Axenya doesn't route AI outputs directly to doctors. Every clinical inference the model surfaces goes to a nurse first, with an explicit reasoning chain: this is what we observed, this is why it matters, these are the papers that support this inference. If the model can't produce that chain, the inference doesn't get surfaced. The nurse reviews and may escalate to a physician. The doctor makes the final decision.
This architecture exists for two reasons. First, a Harvard/Stanford study published recently found error rates in medical AI models remain meaningfully high - not catastrophically, but high enough to warrant human review. Human doctors also make mistakes; the goal is to combine the computational coverage of AI with the judgment and contextual awareness of a clinician. Second, the liability question: AI as a decision-support layer doesn't shift clinical responsibility from the physician. AI as a direct decision-maker does, and neither the regulatory environment nor patient trust supports that transition yet.
The long-term vision isn't replacing the doctor. It's routing patients more efficiently - getting the right person to the right specialist before they've wasted months in the wrong queue - and freeing physician time from the diagnostic work that can be automated, so it can be applied to the cases that genuinely require human judgment.
Frameworks from This Episode
- Outcome-Aligned Revenue Model - Get paid on the measurable delta between your client's healthcare cost growth and the market baseline. No outcome, no fee. Aligns incentives throughout the value chain and makes the technology accountable to the result it claims to produce.
- AI Copilot Architecture - Every AI output surfaces with a transparent reasoning chain and supporting citations before reaching a human decision-maker. If the model can't explain why, the inference doesn't get surfaced. Human judgment is preserved at the decision layer; AI covers the computational breadth no human team could match.
- Layered Continuous Monitoring - No single data source is sufficient for population health inference. Combine financial claims data, wearables, face scan biomarkers, labs, questionnaires, and partner integrations into a unified data lake. Each layer compensates for the gaps in the others; the combination is far more predictive than any individual source.
Tools Mentioned
- Axenya - AI-powered operating system for corporate health. Monitors 100K+ people, runs 95M clinical inferences/month, covers 40+ conditions, and charges on outcome delta rather than service delivery.
- Whoop - High-end wearable health monitor. Referenced as an example of wearable-plus-biomarker integration (combining continuous biometric data with blood-based biomarkers) and as a device that can detect cardiac events and alert emergency services.
Glossary
- Clinical Inference - An AI-generated probabilistic assessment of a patient's health status or risk level based on multiple data inputs. Not a diagnosis - a statistically grounded estimate of likelihood ('the probability of this person having hypertension is above average') that gets routed to a clinician for review. Axenya generates 95 million of these per month across its population.
- Intelligent Navigation - The routing of patients to the appropriate care setting, specialist, or intervention before they self-navigate - which typically means wandering through several wrong providers before finding the right one. Intelligent navigation reduces both cost and time-to-correct-diagnosis by predicting where a patient needs to go, not just alerting that something is wrong.
- Outcome-Aligned Revenue - A payment structure in which a health technology vendor earns revenue only when their intervention produces a measurable improvement in the client's health outcomes or cost trajectory. The opposite of fee-for-service or per-seat SaaS models, where the vendor is paid regardless of result. Creates structural accountability: the vendor's financial interest and the client's health interest point in the same direction.
- Continuous Biomonitoring - Measuring health variables throughout a patient's daily life rather than episodically at clinical visits. Blood pressure at the doctor's office reflects one moment; continuous monitoring captures how it varies during sleep, stress, exercise, and daily activities. Chronic disease management requires this longitudinal picture; point-in-time measurements miss the patterns that predict future events.
- Transparent AI Reasoning - The requirement that an AI system explain its output - what it observed, what that implies, and what evidence supports the inference - before that output reaches a human decision-maker. Axenya's standard: if the model can't produce a reasoning chain with supporting literature, the inference doesn't surface. Designed to address the dual problems of clinician liability and AI error rates.
- Prevention vs. Sick Care - The distinction between healthcare systems designed to detect and prevent disease before it presents symptoms (prevention) and systems designed to treat illness that has already declared itself (sick care). Most modern healthcare systems are optimized for sick care because that's where the incentive structures point. Chronic disease - which has almost no early symptoms - is structurally mismatched with sick care and requires a prevention architecture to address effectively.
- Population Health Management - The practice of managing health outcomes across a defined group (a company's employees, an insurer's members) rather than treating individual patients in isolation. Requires aggregate data analysis, risk stratification, and proactive outreach to the highest-risk cohort before they present clinically. The 5% of a population that drives the majority of healthcare costs is the primary target; continuous monitoring identifies that cohort before they become acute cases.
Q&A
Why is healthcare so hard to transform with technology despite decades of promises?
Two interlocking problems. First, data: health data is fragmented across insurers, providers, labs, pharmacies, and devices - collecting it continuously and at scale wasn't possible until recently. Second, incentives: the current system doesn't financially reward prevention. An insurer who funds a weight-loss program today is unlikely to hold the policy when that investment pays off five years from now. The technological capability and the economic incentive to use it weren't present simultaneously until now. Axenya's thesis is that continuous data collection and AI capable of analyzing it at population scale have finally arrived together - which is why the transformation that Marc Andreessen predicted 12 years ago is only now becoming operational.
What does a 30-second face scan actually measure?
Axenya uses an MIT-developed model that measures the color changes in facial blood vessels across the video frames. The model infers 32 clinical variables from this - including blood glucose, cholesterol, and blood pressure. These are not diagnoses; they're statistically grounded probability estimates. The output goes to a nurse, not directly to a patient or doctor, and the nurse routes it forward if the signal is significant. The purpose isn't to replace a lab draw - it's to flag people who should get a lab draw but wouldn't otherwise seek care, because they have no symptoms that would motivate them to go.
Why does Axenya charge on outcome delta rather than a SaaS fee?
Because incentive alignment is the core problem in healthcare, not just a pricing strategy. Mariano's view: any health technology that gets paid regardless of whether it works is operating under the same misalignment that keeps the status quo broken. Axenya's model - a fee based on how much the client's healthcare cost growth deviates from the market baseline - makes the company's financial interest identical to the client's health interest. 95% of clients have been saving money for three consecutive years. The model also makes the pitch simpler: the technology is free unless it works, at which point you pay a share of the savings it generated.
How does Axenya handle the liability and trust questions around AI medical output?
The architecture keeps humans in the decision loop at every critical point. AI inferences go to a nurse first, accompanied by an explicit reasoning chain: here's what we observed, here's why it matters, here's the literature supporting this inference. If the model can't produce that chain, the inference doesn't get surfaced at all. Nurses escalate to doctors where indicated; doctors make final clinical decisions. This isn't a regulatory workaround - Mariano believes the models genuinely aren't accurate enough yet to operate without human review. A recent Harvard/Stanford study found medical AI error rates remain meaningfully high. Humans also err; the goal is to combine AI's computational coverage with human clinical judgment, not replace one with the other.
What's the business idea in healthcare that Mariano hasn't acted on?
The audit and fraud detection layer between payers and hospitals. Approximately 25% of healthcare expenses are fraud, waste, or abuse - and payers currently have almost no tooling to detect it at scale. AI is well-suited to identify billing anomalies, flag outlier treatment patterns, and compare efficiency across providers. Mariano sees this as a large, underserved opportunity that Axenya isn't pursuing but that someone should be.
What's unsettling to Mariano about AI?
Not the technology itself - the governance gap around it. The mechanisms by which training data is collected (often from people who haven't consented or been compensated), the concentration of economic value in the hands of a small number of AI developers, and the absence of pro-society regulatory frameworks. He believes AI will produce significant wealth concentration, which he sees as a long-run social problem. The technology may be transformative; whether it's deployed in ways that distribute that transformation broadly depends on governance structures that don't yet exist.