How Generative AI Handles Chart-Based Patient Data
One of the trickiest things I ran into while testing AI prompts in a medical dashboard was getting them to understand lab result tables. Almost every EHR (Electronic Health Record) system formats lab values like a spreadsheet, meaning you have rows labeled “WBC,” “RBC,” or “Hemoglobin,” and columns for reference ranges, patient values, and timestamps. But if you paste that into a language model prompt with no structure — it often goes off the rails.
Thank you for reading this post, don't forget to subscribe!
For example, I fed a GPT-4 prompt: “Summarize this CBC (Complete Blood Count) panel and highlight abnormalities.” Then I copy-pasted a text-formatted table from a PDF chart. Even though it looked fine to me, the summaries it generated often ignored the reference ranges or, weirder, flagged normal values as dangerous because it couldn’t correctly match the unit columns.
The fix? I started adding context before the table. Literally one line can help: “Below is a CBC panel with standard US-based reference values. Units are stated next to results.” After tacking that on, the AI stopped mixing up columns and actually flagged the low hemoglobin correctly.
Parameter
Patient Value
Reference Range
Status (AI Generated)
WBC
5.2
4.5 – 11
Normal
Hemoglobin
10.7
13.5 – 17.5
Low
Platelets
180
150 – 400
Normal
This also happens when: the column headers are missing, or if units are mixed (like “mg/dL” and “mmol/L”). AI models can hallucinate and assign completely wrong ranges from memory. Adding units in every cell helps, even though it clutters the prompt.
Two better structures to try:
JSON format: Listing the results as a JSON object makes language models behave better. Example:
It produced summaries that used consistent medical language and even suggested retesting timelines.
Bullet list format: For models that don’t handle JSON well, label each value in a list:
Hemoglobin: 10.7 g/dL (normal range 13.5 to 17.5)
WBC: 5.2 x10^9/L (normal range 4.5 to 11)
This avoids confusion and usually leads to cleaner decisions.
To wrap up, generative AI can interpret structured lab data accurately — if you give it gentle signposts like format notes, consistent units, or JSON layouts.
Using AI to Simulate Bedside Diagnostic Reasoning
This is where things get spooky good. I’ve been running prompt experiments that simulate medical students doing real-time clinical reasoning: stuff like “A 60-year-old presents with shortness of breath, bilateral rales, and elevated BNP. Differential?” At first, I used ChatGPT to just generate differential diagnoses linearly. But after a few tries, I layered in reasoning steps: “Walk through pathophysiology using cardiac, renal, and pulmonary systems.”
Now the AI stops guessing and actually walks me through why it’s leaning toward heart failure (“elevated BNP, chronic hypertension, chest X-ray findings…”). It’s not perfect — sometimes it claims pleural effusions would increase lung sounds, which isn’t right — but overall, it mimics a verbal attending-level breakdown.
Here’s how I structured the prompt that worked best:
“Patient: 60 y/o M, SOB for 3 days, bilateral crackles, elevated BNP. Provide step-by-step diagnostic reasoning considering cardiac vs. pulmonary causes. Conclude with most likely diagnosis and next step.”
That extra instruction to walk through steps makes a huge difference. If you just say “What’s the diagnosis?” — you get a blind guess. When you force structure, you get reflection. This mirrors how real clinicians are trained.
If the model gets off-track? Add a failsafe by specifying: “Use only findings mentioned. Do not invent labs or symptoms.” That helped a lot when the model tried to assume things like fever or edema that weren’t in the case writeup.
Where this goes wrong:
It can sometimes default to common diagnoses even when data suggests rare conditions — just like some humans.
If the age, sex, or chief complaint is vague, the model may float in circles.
Ultimately, by nudging generative AI to think in steps — not just answers — you get output that models bedside reasoning surprisingly well.
Prompting For Radiology Report Summaries
I tested around 20 CT thorax reports with AI using direct copy-paste into prompts like: “Summarize the findings in plain English suitable for a nurse handover.” The biggest issue: pronoun confusion and jargon retention.
The model nailed concepts like “no pulmonary embolism,” but often retained phrases like “right-sided basilar atelectasis” — which isn’t helpful unless the receiver has a medical background.
So I sliced that down by appending: “Avoid Latin or specialized terminology. Assume the reader is a general nurse.” That cleaned up the summary to something like:
“There are no clots in the lungs. Some mild areas of partial lung collapse at the base, likely due to shallow breathing.”
Way more useful.
When it absolutely fails: If you drop more than two reports into one prompt, it blends them. You’ll get hybrid monsters like “mild right pleural effusion in Patient A” when B had it. The fix: one report per prompt session. Or, separate clearly with labels and delimiters like ###.
Better outcomes: Ask it for “bullet points” or split summaries: • Diagnostic information • Follow-up suggestions • Unclear findings This makes it format thoughts better instead of scribbling long prose.
The bottom line is: with radiology reports, less is more — one report, one transformation, one prompt.
Quick AI Prompts for Triaging Symptoms
This one is trickier than it sounds. I thought prompting: “Triage this: 29M, chest pain, intensity 7/10, radiates to left arm, onset during exercise” would spit out something like “emergent, possible MI (myocardial infarction).” Instead, depending on phrasing, it sometimes downplayed urgency (!).
What worked better: Asking for weighted triage decisions by risk factor:
“Based on age, sex, onset, and pain features, rate this symptom as low/medium/high urgency for further cardiac testing.”
That prompt worked because it asked for graded reasoning, not just a yes/no. It returned: “High urgency. Male under 40 with exertional chest pain and radiation is concerning for underlying cardiac cause despite lacking classic risk factors.”
Prevent confusion: Clarify if vitals are normal or missing. The AI sometimes assumes low BP or tachycardia — it projects context that isn’t supplied otherwise. Add a line like: “Vitals not recorded. Patient alert and oriented.” That locks it down. 👍
Also useful: Let it return more than triage level — specify goal:
“Is it safe to defer this 12 hours?”
“Should this patient receive ECG immediately on arrival?”
Then you’ll get actions, not just labels.
To conclude, you’ll get the most clinically appropriate urgency cues when you structure prompts with decisions in mind, not diagnoses.
Challenges Interpreting Free-Text Nurse Notes
This was such a pain — nurse notes are goldmines for symptom evolution, but they’re also messy, abbreviated, and inconsistent (“Pt AoOx3, c/o cp, denies sob, HR ↑ to 110”). Without support, LLMs get lost in the shorthand. The first few times I pasted nursing logs into GPT-4, the output said stuff like “The patient is confused” — completely misreading “A/O x3” (which means alert and oriented to person, place, and time).
Fix? At the top of the prompt, add: “Note: A/O x3 = alert and oriented to all spheres. c/o = complains of. HR = heart rate.” Or even better: “Use medical abbreviations defined below. If unsure, do not guess meaning.” That made a huge accuracy jump. No more hallucinated mental status changes.
Another trick: Use this style — all in one input:
---
Nurse Note:
Pt AoOx3. VS stable. C/o chest pain. HR increased from 82 to 110 at 0800. No SOB.
---
Summarize symptom progression. Include trends and any cardiac flags.
That structure helped the model find temporal trends — it called out the rising HR on its own. If you leave the data floating unstructured, it often just rewrites in fancier words without interpreting.
As a final point, free-text input only works when you handle abbreviations and supply structured goals — otherwise it’s just noise to the AI.
