Cadence banks $100M for the boring part 🩺, AI reads your EKG's secrets 🫀, AI for low yield labs 🧪

The $100M bet isn’t on a smarter doctor. It’s on the space between visits.

Cadence raised a $100M Series C led by Spark Capital, valuing it at $1.23B, to scale “regulated AI agents” that manage chronic disease for older adults.

It runs across 20+ health systems and 100,000+ active patients, and claims peer-reviewed outcomes plus ~$2.7M/week in Medicare savings.

A safety-net PCP put the real problem in one line this week: “care is episodic, illness is continuous.” Most clinical risk lives in the unwatched space between visits.

The durable money in clinical AI isn’t chasing the model that answers your question best. It’s chasing whoever owns the watching between visits.

That reframes the whole stack. The moat isn’t the model — models are commoditizing by the month — it’s the value-based contract, the health-system integration, and a number like “$2.7M a week” that a CFO can underwrite.

But agents that manage a chronic patient between visits also break the thing malpractice insurance rests on. One flawed rule propagates a single error through thousands of patients at once — and no carrier wrote a policy for that.

😤 “This is just RPM with a chatbot bolted on.” No. RPM was a billing code waiting for a Bluetooth cuff. This is a company taking risk inside value-based contracts and showing peer-reviewed outcomes. The difference between “a device that reads a number” and “an agent that’s accountable for the number” is the whole company.

😤 “Agents managing chronic disease between visits is a lawsuit waiting to happen.” You’re not wrong. That’s the unpriced part.

😤 “$1.23B for chronic care management? CMS already pays for that.” It does — CCM and RPM codes already exist. That’s exactly why it’s investable: the reimbursement rail is laid, and the agent is just the thing that finally runs the volume the codes always assumed.

💡 80/20: The buildable insight isn’t “automate chronic care.” It’s that the unglamorous between-visit workflows — titration follow-up, abnormal-result chase, missed-refill outreach — are where agents are actually shipping. Prototype one on synthetic Synthea patients and you’ve built the exact shape the money is funding.

---

📡 Builder’s Radar

A practicing cardiologist just landed the first FDA clearance for multi-condition cardiac AI — and put it inside OpenEvidence

Dr. Pierre Elias — a general cardiologist at Columbia/NewYork-Presbyterian — built EchoNext, which reads a standard 12-lead EKG and flags six forms of structural heart disease that usually stay hidden until a patient is symptomatic.

It was trained on more than 700,000 ECG-echo pairs and outperformed cardiologists, including cardiologists using AI. The new distribution deal pushes it into OpenEvidence, used by half of US clinicians.

The EKG you already order on nearly everyone just quietly became a structural-heart-disease screen. And the person who built it still sees patients.

😤 “Screening AI on every EKG is an incidentaloma factory — you’ll drown cardiology in downstream echoes.” Fair, and the false-positive math is the whole ballgame. But this isn’t flagging the worried-well; it’s surfacing valve disease and infiltrative cardiomyopathy in people already getting an EKG for another reason. The question to ask the vendor isn’t sensitivity — it’s positive predictive value in your population.

💡 80/20: The signal here is that the most defensible clinical AI is built by clinicians on data they already understand. If you have an underused data stream in your specialty (the EKG, the CTG, the spirometry curve), that’s a moat hiding in plain sight.

---

The FDA dropped its enforcement action against Whoop after the company tweaked its blood-pressure feature — an early read on how the agency’s deregulated, move-fast posture toward consumer wearables actually plays out.

States are surging ahead on healthcare AI regulation while the federal government deregulates, especially around AI in prior auth and coverage decisions — meaning the compliance map for builders just fractured into 50 pieces.

---

People doing things:

Farhang Dehzad (Founder, Omi Health) — open-sourced a safety eval of 8 frontier models on 2,400 generated SOAP notes: 12 hallucinations vs. 520 omitted safety-relevant facts. AI scribes don’t fabricate so much as they quietly forget — and the safety conversation is over-indexed on the wrong failure mode.

Sergei Polevikov dissects Doximity’s new “Ask” white paper as a de-facto answer to last week’s Nature Medicine paper — and to OpenEvidence’s combative response. His point lands: a fine-tuned RAG system marketed as “unlike an LLM” is still an LLM, and “we RAG’d it” is weakening as a moat.

---

🛠️ From the Workbench

SmartAlert — ML-driven CDS that cuts low-yield inpatient labs (NEJM AI, Jun 23)

A real-world implementation study, not a benchmark exercise: an ML model that flags repetitive, low-yield inpatient lab orders and nudges clinicians to stop, reducing waste inside the EHR.

The builder lesson is the unsexy one — the most shippable clinical AI right now isn’t diagnosis, it’s utilization. A model that quietly removes a redundant daily CBC clears governance faster than one that proposes a diagnosis, because nobody dies from the AI declining to draw blood.

💡 80/20: Pick one over-ordered test in your shop, build the “do you really need this?” rule against synthetic data, and measure your own override rate before you ever pitch it.

---

🎙️ From the Pods

🎙️ Radio Advisory — “Boom, bust, or bubble? Rock Health weighs in on digital health funding in 2026”

The pearl: the moat is increasingly not AI, because AI is being commoditized by the month — and last year ~40% of digital-health venture dollars went to mega-deals over $100M, a barbell of haves and have-nots. The bigger shift is LLMs becoming the consumer “front door” for care, quietly disintermediating primary care.

🔇 Speaker Blindspot: Survivorship bias — the named proof that “AI-native compresses sales cycles” is OpenEvidence, Hippocratic, Commure, Abridge: the four companies that won the mega-deal lottery. The “have-nots” she flagged one sentence earlier never get named, so the evidence for the thesis is quietly drawn only from the survivors.

---

🎙️ Lifers — “Joanna Strober, Midi Health”

The pearl: “B2B for show, B2C for dough.” Strober’s read is that nobody shops for “primary care” — they shop for a specific problem (menopause), so the moat she built was a trusted consumer brand, not a feature set. People come back to a brand; they don’t come back to a portal.

🔇 Speaker Blindspot: False consensus from a single win — the same DTC-brand logic that worked for menopause failed her in childhood obesity, where the population was Medicaid and there was no consumer wallet to sell to. She names this, but doesn’t fully reckon with the implication: the playbook generalizes to who can pay, not to who’s sick. The hardest patients are exactly the ones the B2C model routes around.

---

📅 Upcoming: AHIP’s Prior Authorization Modernized: From Commitment to Execution webinar is today (Wed, Jun 24).

---

What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.

— Kevin