Sameed Khan

Skimmed abstract, scrolled straight to appendix - commit / API column is misleading; most of these are late 2024 release models. Only one that is slightly current is Qwen3-30B-A3B-Instruct-2507 (July 2025). This is a completely unrepresentative view of current AI capabilities.

Also honored to show work w/ longtime mentors and colleagues Chad Klochko MD and Steve Soliman DO on novel work using radiomics on muscle ultrasound to predict insulin resistance! We believe that this could expand to less invasive and easier-to-access screening for metabolic disorders (instead of a needle now you get POCUS) We show that you can take a simple logistic regression classifier, train it on radiomics from deltoid US in a messy retrospective dataset to predict T2DM and that same model will predict insulin resistance very well in patients from a completely different institution! #RSNA25

Had the incredible opportunity to present work w @cshahrad at #RSNA25! We evaluated the real-world performance of an automated ASPECTS scoring tool (RapidAI, iSchemaView) on one year of stroke alert NCCTs at Cleveland Clinic and found that AI routinely overcalls ischemia, especially on LVO negative cases which, while outside the validated use case for many AI ASPECTS tools, is far more common than LVO positive cases.

Been working on AI dev for a while but finally built my first web app - 10 weeks, probably > 150 hours and 5000 lines of code: a simple localhost interface to annotate images via keyboard a la Anki.

Who knew some CRUD could make a man so happy?

Me on rounds checking for labs updates so I can maximize my limited usefulness

Did I do the data labeling I was supposed to do this weekend? no Did I instead just spend all that time to make a TOOL to do the labeling? yes

Claude Opus extremely cracked and >>> GPT-4. Went from 0 to this web app skeleton within one day (very basic Anki-style radiology report annotator tool). Little to no HTML/CSS knowledge, very novice Rust knowledge, never used Rocket before, little to no web dev experience.

Getting Rust-y -

Current state of AI

Was it used as part of a study? Nothing What were the collection methods? Nope Were these patients even consented? Idk Who even is the listed author - "John?" No idea, has no Kaggle activity other than dropping a massively cited dataset and disappearing

You'll find a lot of citations for the Frankfurt Hospital diabetes dataset, but they all point back to Kaggle, where you find... Nothing lol

My favorite study: "Towards a Smart Healthcare System: An Architecture Based on IoT, Blockchain, and Fog Computing" Collect wearable health data --> collect on the blockchainzzz--> do magic ML stuff --> deliver result to patient and patient care team. 96.4% prediction accuracy!

If you do a quick Google search for "diabetes machine learning prediction frankfurt dataset" here are some of the top results: