Clinical AI & Healthcare
Evidence, systems and care.
Clinical AI, agents, and evidence tools, with strategy, operations, finance, and health systems.
11 notes
The price is not the cost: how healthcare turns care into a payment
Healthcare does not have one price. A charge, a negotiated allowed amount, a patient's share, an insurer's payment and a provider's net revenue are different numbers. Understanding the path between them is essential before comparing hospitals, contracts or countries.
When hospitals merge: value, synergy and bargaining power are not the same thing
A hospital acquisition can create operating value, transfer value through stronger bargaining power, or do both. Reading the deal from the seller, buyer, patient and payer perspectives reveals what a single purchase price conceals.
Five statements, one hospital: reading healthcare finance as a system
A hospital's financial statements are not five separate reports. Together they explain what the organization owns, what it owes, whether care operations sustain themselves, where cash moves and which restrictions or risks the headline numbers conceal.
No margin, no mission? How to read hospital financial ratios without being misled
Hospital ratios are useful questions, not automatic verdicts. This guide connects profitability, liquidity, capital structure and asset management, then uses Cleveland Clinic's 2021 audited statements to show how operating and total margin can tell different but simultaneously correct stories.
Building AI products for health and fitness: autonomy must be earned
The opportunity in healthcare AI is not to automate the clinician first. It is to define one valuable capability, evaluate it with real cases, learn from its failures and expand autonomy only when evidence justifies it.
Healthcare management is the work of aligning four systems
A healthcare organization does not improve because every department works harder. It improves when mission, clinical care, operations and finance point in the same direction. Saudi Arabia and Colombia organize those forces differently, but the manager's real job is the same: turn resources into a patient journey that is accessible, safe and continuous.
Grounding an LLM in evidence: faithfulness, citations and hallucinations
Retrieval puts the right passages in front of the model, but the model can still drift and state something the passages never said. Grounding is the discipline that fights that: cite every claim, forbid mixing chunks, measure faithfulness. In any high-stakes domain, a confident unsupported answer is the failure to engineer against.
Query understanding: from patient language to clinical search
A patient says 'my shoulder went numb after the dog pulled the leash'. A clinical corpus is written in terms like 'traction injury with paresthesia'. Three techniques bridge that gap before retrieval even runs: query rewriting, entity extraction, and hypothetical-document embeddings.
Opening the black box: SHAP and permutation importance
A powerful model can predict a risk score and still be useless to a coach or clinician if it cannot say why. Two model-agnostic tools fix that: permutation importance for what matters overall, and SHAP for why this one case scored the way it did. Both work without retraining.
The emerging architecture of AI health coaching
The 2024 to 2026 research keeps landing on the same loop: perceive, infer the person's state, generate options, select a strategy, communicate without prescribing, observe, update. The new idea is not a smarter chatbot. It's separating the policy that chooses what to do from the text that says it.
What Stanford's GPTCoach & Bloom actually proved
Stanford's coaching agents showed that an LLM can follow non-prescriptive, motivational-interviewing strategies and adapt to a person's stage of change. The most honest result is humbling: in a 4-week trial, LLM coaching did not beat the control group on activity. Its advantage was psychological, in beliefs, enjoyment and self-compassion.