Clinician’s Guide to AI: Models, Data & Better Decisions

Learn how large language models actually work, and how the data you give them changes their answers from generic advice about a population into grounded insight about your patient.

This session is the first in a new Heads Up series on AI education for clinicians. Tim Oates, PhD, is Chief Data Scientist at Synaptiq and a Professor of Computer Science. He starts by explaining how LLMs work under the hood, then moves through a “context ladder.” At the bottom, a model knows only what it learned in training. At the top, it’s connected to charts, labs and wearable streams, and it works between visits. Using a synthetic perimenopause case, Tim shows how each added layer of data changes what the model can reason about, where hallucinations come from, how to write a prompt that works like a real consult request, and why human oversight isn’t optional. Heads Up founder and CEO Dave Korsunsky explains what it takes to connect AI to EHR, PDF and wearable data, and how a practice can personalize that “last mile” to the way it works.

What you’ll learn

What an LLM actually is, and why context is everything

A language model predicts the next word from probabilities. A “large” one does the same thing, but it’s trained on nearly all written text and has trillions of parameters. Once you understand that, it’s clear why a little extra context can move an answer from plausible to precise.

Why hallucinations happen, and what never to trust at “rung zero”

A made-up citation looks real because it’s statistically likely. Tim’s rule: use a bare model for explanations, trade-offs and drafting, but never trust its numbers, citations or doses without checking them.

How to write a prompt like a consult request

Name the patient, ask a specific question, say what data the model should use, limit the length and format, require evidence for every conclusion, and ask what information would change its answer.

What happens when AI can see the chart, and the wearables

Connected to labs, medications, portal messages and months of HRV and sleep data, the model stops giving general advice. It starts catching things like a missed repeat draw, a medication the patient stopped taking, or a decline that began two months before the patient noticed it.

AI between visits, and where to draw the line

Background agents can write pre-visit summaries, watch real-time feeds and rank what needs your attention. They should never change a plan, cancel an order or tell a patient what to do.

Summary

It’s read everything, and it’s met no one

Tim starts by demystifying the technology. A language model only predicts the next likely word. He trained a small model on 2.2 million recipes to show the effect of context: when it looks back only one word, it writes nonsense like “applesauce or more for frying chicken stock.” When it looks back two words, the output becomes coherent. Frontier models apply the same principle to almost everything ever written, and transformers let them consider the entire context window. The model stays the same in every step of the talk. Only the information you give it changes.

Rung zero: powerful, but never trust a specific

A frontier model has read more medicine than any clinician. That makes it excellent for explaining mechanisms, comparing treatment options, translating content, changing reading level and drafting. It also invents things convincingly. Tim shows a citation that looks perfect but doesn’t exist: the journal, year, volume and page count are all plausible. His rule is to trust the shape of the explanation, never the specifics, and to check every number, citation and dose. Agentic tools that can search the web help, but they don’t fix the problem completely.

Write the prompt like a consult request

A vague prompt such as “check this patient out” gets a vague answer about nobody in particular. Tim rewrites a synthetic perimenopause case as a structured consult. It tells the model to use only the case provided, asks for the three most likely contributors ranked by likelihood, requires supporting evidence from the case for each one, asks what additional information would change the answer, sets a 200-word limit, and excludes treatment recommendations. He also recommends treating the model like a colleague in a Socratic dialogue, building shared context back and forth. And he warns that clear reasoning is not the same as evidence.

Stop typing, hand it the chart

Connected to real patient data, the model moves from general advice to specific findings. In Tim’s synthetic example, it notices that a repeat ferritin was ordered but never drawn, that the patient stopped her iron because of GI upset although her medication list still shows it (she said so in a portal message), and that her A1c is rising. Dave explains the engineering behind this: a secure connection between the LLM and the EHR, and extraction of data locked in PDF lab reports into clean, structured tables. The caveat is that the chart is not the patient. If something was never recorded, the model can’t see it.

Wearables are a movie, not a photograph

A lab result is a snapshot. Wearables produce thousands of data points per patient per day, far too many to paste into a prompt. The solution is to summarize with a purpose, using nightly averages, rolling averages or time above a threshold. Tim points out that every summary choice builds in a bias about what matters. In the synthetic case, HRV and sleep efficiency start falling in June while steps and weight stay flat, two months before the patient reports feeling worse in August. Continuous data can show a change before the patient does.

Between visits: the curated morning list

At the top of the ladder, AI works in the background. It can write pre-visit summaries, watch real-time feeds and flag emerging trends. The hard part is deciding what’s worth interrupting you for, since alert fatigue is a real risk. Ranking helps: show the top five, and if those are useful, show the next five. Dave describes where Heads Up is heading: agents that run around the clock across every connected device in a practice, combine hard-coded thresholds with room for the model to find clinically relevant patterns, and deliver a curated list of who needs attention each morning.

The last mile: your practice, your way

Tim closes with a warning: there’s a big gap between an impressive demo and a tool you can rely on every day. Human oversight is required, and you must always be able to trace a conclusion back to its evidence. Dave explains why frontier chat tools aren’t systems of record. They have no database, no structured patient record, no integrations, no patient portal and no way to customize them to how a clinic practices. Heads Up provides that layer and personalizes it to each practice, including what the pre-visit report looks like, how chart notes are drafted and which signals matter. In one recent project, an agent drew on more than 100 scientific papers, the clinic’s full peptide and supplement protocol library, and 10 diagnostic reports per patient, and cut treatment plan generation from two hours to twenty minutes.